2026-11-08

15 repos · 0 papers · 58 scholar

Showing 73 of 73 items

GitHuboptimization

cinebench

Cinebench 2026 pre-configured environment for professional desktop software utility configuration.

3000
GitHubaccelerators

tiny-NPU

opensource NPU for LLM inference (this run gpt2)

SystemVerilog29670
GitHubframeworks

MangaJaNaiGUI

GUI for upscaling manga and images with ONNX models, using TensorRT, TensorRT-RTX or DirectML

C#26716
GitHuboptimization

cinebench-desktop

Cinebench 2026 Standalone Desktop Setup with Pre-Activated Features

HTML2440
GitHubframeworks

inference

High-performance Ultralytics YOLO inference in Rust with ONNX Runtime, GPU backends, CLI, and WebGPU/WASM.

Rust21731
GitHubmodel-compression

Tator

Self-hosted dataset management and model-assisted annotation, with multi-user review, quality analysis and dataset pruning.

Python9213
GitHubai-hardware

Spiking_GNN

Bio-inspired spiking graph neural network with neuromorphic learning and connectome-inspired modeling.

Python10
GitHubai-hardware

genx320-person-detection

Live person detection with a Prophesee GenX320 event camera on an STM32F746 (Cortex-M7): firmware, recording, curation, simulation, training and evaluation tools of the N-People dataset

C10
GitHubai-hardware

IHP__SoC6169

Open-source 32×32 XPCAM-based in-memory computing architecture for binary neural networks, featuring reconfigurable time-domain sensing and XNOR-popcount operations in IHP SG13CMOS5L technology.

Verilog00
GitHubai-hardware

HopOS

A bare-metal Go operating system for edge computing. No Linux — one static Go binary is the OS: apps run natively on their own CPU core in their own hardware-enforced memory partition.

Rust00
GitHubai-hardware

synapse-spike

Neuromorphic Spiking Neural Network (SNN) & Synaptic Plasticity Dynamics Simulator

Python00
GitHubai-hardware

NeuroShaqma

小沙大脑:288,964 神经元脉冲皮层(果蝇 2.07x/4.16x),12 岁初中生在家用 2080Ti+4090 做的 SNN,主力开发 DeepSeek-V4-Pro

Python00
Scholarai-hardware

Efficient and Green Service Function Chain Deployment and Scheduling for Mobile Edge Computing

Service Function Chain (SFC) decomposes a user’s overall service into multiple service components and links them in an orderly manner, providing more flexible, efficient, and responsive services for compute-intensive applications in mobile edge computing. Existing research primarily focuses on optimizing the deployment cost and system efficiency of SFC. However, the dynamic nature of user trajectories, along with the absence of adaptive scheduling and service component sharing in complex and heterogeneous environments, hinders existing approaches from achieving optimal resource efficiency. To achieve adaptive optimization of SFC deployment and efficient component sharing across multiple SFCs, we propose an SFC Deployment and Scheduling strategy based on the Soft Actor-Critic (SAC) algorithm, named SAC-DS. The proposed SAC-DS strategy first employs a Long Short-Term Memory (LSTM) model to predict user trajectories, thereby accurately forecasting future service demands. Based on these predictions, an SAC-based agent is trained to dynamically optimize SFC deployment and scheduling policies in real time. SAC-DS integrates SFC deployment, component sharing, and scheduling into a unified framework, effectively improving system throughput while minimizing overall deployment costs, thereby enhancing user experience. Compared with existing baseline methods, SAC-DS improves network throughput by 3.1%–18.1% and reduces average cost by 1.8%–10.0%.

Xue-Jian Chi +71IEEE Transactions on Mobile Computing
Scholarai-hardware

Resource-Efficient and Robust Collaborative Computing in Satellite-Terrestrial Networks

With the rapid proliferation of large Low Earth Orbit (LEO) satellite constellations, an enormous volume of in-orbit data is continuously generated and must be transmitted to the ground for processing. However, traditional LEO satellites downlink raw data to the ground with severely restricted transmission capability, creating a major communication bottleneck. Satellite computing, which leverages onboard processing capacities to analyze the raw data in orbit, has emerged as a promising solution to alleviate the downlink pressure. However, this approach shifts the bottleneck to inelastic computational and energy resources. Two key challenges remain largely unaddressed in existing satellite systems. First, onboard computation cannot keep pace with the high volume of captured images. Second, the energy supply available for satellite operations is strictly limited. Moreover, model deployment in space remains hindered by Single Event Upsets (SEUs), caused by high-energy particles that can flip memory-stored model bits and compromise reliability. These constraints make it essential to jointly optimize onboard computation, energy usage, SEU robustness, and data transmission to enable effective satellite-ground collaboration. In this work, we exploit the limited onboard computation and communication resources to enable satellite-ground collaboration and propose a synergy framework, TargetFuse, for resource-efficient object counting. Building upon TargetFuse, we develop TargetFuse+, which integrates SEU-resilient model protection and bandwidth- and energy-aware data throttling to minimize counting errors under strict resource constraints. We conduct comprehensive evaluations using parameters and real-world data from the BUPT-1 satellite. Extensive experiments demonstrate that TargetFuse reduces counting errors and improves bandwidth efficiency compared to the vanilla baseline, whereas TargetFuse+ preserves inference accuracy under SEUs while maximizing downlink utility.

Qi-Yang Zhang +81IEEE Transactions on Mobile Computing
Scholarhardware-design

A Dynamic Compact Model for Phase-Change Memory With Coupled Thermal--Electrical--Phase Dynamics

A dynamic thermal–electrical–phase (TEP) compact model for phase-change memory (PCM) is presented and implemented in Verilog-A for device-to-subarray simulation. The model establishes a closed-loop coupling mechanism in which terminal electrical power induces self-heating, temperature governs phase dynamics, and the resulting phase configuration in turn modulates electrical conduction and threshold-switching behavior. Effective latent heat, quench-dependent solidification, and distinct threshold and holding conditions are formulated using continuous and bounded state equations. Calibrated against measurements from a 40-nm 64-Mb prototype chip, the model accurately reproduces the I–V characteristics, bidirectional programming, and multilevel programming capability. Monte Carlo simulations with parameter variations further capture device-to-device variability. Array-level validation demonstrates complete SET–read-- RESET–read operation in 1024 explicitly instantiated cells, while transient simulations of up to 32 768 cells complete without convergence failure. Overall, the model provides a physically consistent, numerically robust, and computationally efficient framework for PCM memory and in-memory-computing circuit design.

Qing-Yu Wu +60IEEE Transactions on Electron Devices
Scholarhardware-design

A Resource-Aware Fixed-Point FPGA Architecture for Distilled Financial Direction Classification

The following sections outline a constraint-driven pipeline for the deployment of financial inference based on FPGA hardware where the design of a standard ML is bound from the onset by implementation limitations. The proposed system is developed with three practical constraints, i.e., only daily OHLCV market data, a single DSP48E1 multiply-accumulate unit utilizing the PYNQ-Z2 (XC7Z020), alongside 8-bit Q1.6 fixed-point arithmetic. As opposed to treating these as separate limitations, the investigation frames them as a co-design flow where hardware constraints guide the system selection, distillation strategy and numeric representation. A Gradient Boosting teacher algorithm is distilled down to a Logistic Regression student which can be performed as a single-pass fixedpoint dot product. The model is subsequently assigned to a sequential MAC datapath via verilog. Training artifacts, quantized software inference, as well as RTL behavior are connected via a bit-exact cross-domain verification chain. AAPL and HDFCBANK are shown without the RTL modification of the pipeline to show that comparable hardware is capable of being reconfigured by just retraining and replacing the parameters. Post-route implementation achieves the 100 MHz timing target employing one DSP48E1 alongside no BRAM, alongside deterministic 0.45 $\mu \mathrm{s}$ inference latency. The principal contribution of this work is a reproducible methodology to translate constrained financial ML models through faithful FPGA inference, focusing on deployment correctness and portability beyond claims of universal stock predictability.

Kevin Anthony +20
Scholarhardware-design

Hazard Detection Sensitivity Metric for Hybrid Assertion and Coverage-Driven Verification of a 5-Stage RISC-V Pipeline

The processing steps in pipelines may be difficult to verify because there are both data hazards and control hazards that can cause incorrect execution time and degradation of performance if not addressed correctly. Traditional directed testing is sometimes inadequate to ensure that all possible hazard scenarios have been run. In this paper, the hybrid verification approach to Assertion-Based Verification (ABV) and Coverage-Driven Verification (CDV) for a 5-stage RV32I RISC-V processor is presented. Hazard behavior is monitored using System Verilog Assertions (SVA) and the verification completeness is measured by functional coverage. A structured mutation campaign with 12 hazard-related RTL mutants is conducted to assess the success of the verification environment. From the mutation results, a new metric called Hazard Detection Sensitivity (HDS) is proposed to measure the detection capability of the verification environment for hazard-related faults. The first mutation campaign resulted in 50.0% HDS with identified gaps in coverage of forward and hazard-handling scenarios. It was then followed by a coverage-guided benchmark augmentation process such that the HDS was raised to 91.67%. In addition, Cycles per Instruction (CPI) analysis was conducted with load-use hazard densities varying, to examine how pipeline hazards affect performance. The results indicate that the proposed framework successfully detects faults and discovers the gaps in verification in the hazard-sensitive processor design with 41.67 percentage point improvement by selectively augmenting the benchmarks.

Ranjan A. R. +30
Scholarhardware-design

Multi-Abstraction Evaluation of Baseline and Optimized 4-Bit Comparator Designs from Custom CMOS to Post-Layout Signoff

Magnitude comparators are essential building blocks in digital systems and are widely used in arithmetic units, control logic, and signal processing applications. Their performance and efficiency are strongly influenced by design abstraction level and optimization strategy. This paper presents a multi-abstraction evaluation of baseline and optimized 4-bit magnitude comparator designs, spanning custom transistor-level design, RTL-level synthesis and static timing analysis, and post-layout signoff validation. At the circuit level, the baseline and optimized designs are implemented using custom CMOS design in Cadence Virtuoso. At the RTL level, the designs are implemented in Verilog, synthesized with Cadence Genus, and analyzed with Cadence Tempus under the same clocking and constraints. The optimized architecture uses logic restructuring, precomputation, and early-termination strategies to minimize redundant switching activity. To further confirm physical realizability, the optimized RTL design is processed through a complete open-source RTL-to-GDS flow with OpenLane, allowing for parasitic-aware post-layout analysis. A fair comparison is made by ensuring the same functionality, register boundaries, and timing constraints are maintained for all designs. Experimental results show that the optimized comparator achieves area and power savings while maintaining setup and hold timing margins, both pre- and post-layout. The optimized 4-bit comparator results in 51.57% faster propagation delay and 55.53% less power consumption than that of the initial design. Even though there is a 10.73% rise in area usage, the slack time has been improved by 22.8%. Post-layout analysis yields a total power usage of $10.1 \mu \mathrm{W}$, no timing errors, and passing DRC and LVS verification.

M. M +20
Scholarhardware-design

Design and Implementation of an Optimized RLS-NR and RLS-GS for Ecg Signal Enhancement Using FPGA

Electrocardiogram (ECG) signals are commonly affected by noise such as baseline wander, power-line interference, and motion artifacts, which reduce diagnostic accuracy. This paper presents the design and FPGA implementation of an optimized Recursive Least Squares (RLS) adaptive filter for real-time ECG signal enhancement. ECG data obtained from the MIT-BIH Arrhythmia Database is preprocessed in Python and transmitted to the FPGA using UART communication. To reduce hardware complexity, Newton-Raphson (NR) and Goldschmidt (GS) reciprocal algorithms are integrated into the RLS architecture to replace conventional division operations with iterative multiplication-based computations. The proposed architecture RLS-NR and RLS-GS are implemented using Verilog HDL in Xilinx Vivado on an Artix-7 FPGA platform. Hardware performance metrics including LUT utilization, flip-flop usage, timing, and power consumption are evaluated. Experimental findings show low power consumption of 0.082 W for the Goldschmidt architecture and 0.083 W for the Newton-Raphson architecture, as well as effective hardware utilization. Real-time UART-based ECG visualization confirms successful signal enhancement and reliable hardware operation. The proposed system provides an efficient solution for FPGA-based biomedical signal processing applications.

Akshitha Goud Dukanam +20
Scholarhardware-design

CORDIC Based Multiply and Accumulate Units Using Lower OR Adders

Approximate computing has become a powerful tool to enhance the area efficiency, power efficiency, and computational latency of error-tolerant applications like image processing, multimedia systems, and machine learning accelerators. In Digital Signal Processing (DSP) and Artificial Intelligence (AI) systems, optimization of arithmetic operations is crucial, and Multiply-Accumulate (MAC) units are the building blocks that can significantly affect overall system performance. Herein, two modified approximate MAC2 architectures using the approximation techniques Reconfigurable Lower-part OR Adder(ReLOA) are proposed MAC 2-LOA 25 and MAC 2-LOA 50. In the proposed architectures, the least significant bits are approximated by OR-gate-based logic with the aim of limiting hardware complexity while ensuring acceptable computation accuracy. 25% of the lower bits are approximated when using MAC2-LOA25, and 50% of the lower bits are approximated when using MAC2-LOA50. Implementation and simulation of the architectures were done with the help of Verilog HDL and Vivado simulator from Xilinx. Comparative analysis is performed on basis of the area, power, delay, and accuracy. The experimental results confirm that the proposed approximate architectures lead to substantial reduction in the hardware resources and delay as compared to the conventional exact MAC architectures, which is desirable for VLSI systems with lower energy consumption.

Gunnala Meghana Sri +10
Scholarhardware-design

Fused SIM(S)D/Scalar Functional Units for RISC-V P-B-Zkn Instruction Set Extensions

Royalty-free, open-source, modular and extensible have catalyzed the wide adoption of the RISC-V (RV) instruction set architecture (ISA). Packed SIMD (Single Instruction, Multiple Data) instructions (P-SIMD) enable accelerating data parallel workloads in AI/ML and bit manipulation instructions (BMIs) cater for crypto, signal processing, etc. The Zk-extension addresses AES, SHA-2, entropy source, and crypto BMIs. Pluggable 4 stage pipelined, 11 fused SIM(S)D/Scalar, processor core EXE stage functional units, for RV-32/64 P-B-Zkn in Verilog to execute 323 P-SIMD, 103 BMIs (v1.0.0+0.93 draft), 11 AES, and 13 SHA2 instructions have been presented here. Design, simulation at the functional level, verification, and performance evaluation have been performed with Vivado, QuestaSim, and UVM. The synthesis has utilized 21,273 LUTs + 4107 FFs for 323 P-SIMD, 23,070 LUTs + 5854 FFs with 103 BMIs/Zbkbcx + Zknh as SIM(S)D on the VCU-108 platform which has Virtex Ultrascale device. Evaluated 5 different merged S-Box schemes for AES for area/latency tradeoff on FPGA and opted for Alexander Maximov proposed based on MUX31/I and 2 I/P gates. The unified RV32/64Zkned scalar unit increased the resources to 24301 LUTs and 6991 FFs. The units were integrated with a 6-stage pipelined RV64IMAFD core with L1-I,D cache, AXI bus I/F etc. and synthesized for FPGA with 158935 LUTs, 71080 FFs, 112 BRAMS, and 25 DSP slices.

Nancy Gupta +10International Symposium on VLSI Design and Test
Scholarhardware-design

Design and Development of a Reliability-Aware Verilog-A LUT Framework for 5 nm Nanosheet SRAM Arrays

This paper presents a TCAD-calibrated, look-uptable-based Verilog-A modeling framework, and uses it to perform a comprehensive reliability analysis of a 4×4 Static Random Access Memory (SRAM) array implemented using Nanosheet Gate-All-Around (GAA) Field-Effect transistor technology, with a focus on device-trap interactions and Bias Temperature Instability (BTI) degradation. Trap-induced current degradation is characterized for both band-tail and deep-level interface traps at a concentration of 1 × 1013 cm-2 using Synopsys TCAD device simulations. The degraded drain current characteristics are incorporated into HSPICE circuit simulations via the TCADcalibrated Verilog-A look-up-table model to evaluate SRAM performance metrics such as Write Delay, Hold Noise Margin (HNM), and Read Noise Margin (RNM) under fresh, Positive BTI (PBTI), Negative BTI (NBTI), and combined degradation scenarios. Results show that deep-level traps cause higher current reduction than band-tail traps at the same concentration. Under combined band-tail PBTI and NBTI stress, write delay increases by 16% and HNM decreases by 5.3% relative to the fresh device, while NBTI alone collapses RNM by 48.9% and reduces HNM by 9.3%. Under deep-level trap degradation, write delay increases by up to 47% relative to the fresh device.

Charanya Bysani +40International Symposium on VLSI Design and Test
Scholarsynthesis-pnr

An Anonymous and Efficient Key Agreement Protocol for Secure Surgeon-Robot Interaction Over Tactile Internet

The emergence of the Tactile Internet has revolutionized remote healthcare, enabling real-time surgical operations through highly responsive and low-latency communication between surgeons and robotic systems. This next-generation technology facilitates real-time haptic feedback and high availability, empowering time-critical applications such as remote surgery. However, the increasing interconnectivity among these entities introduces significant security and privacy challenges, including real-time authentication, protection against physical tampering, and impersonation attacks. To address these concerns, this paper proposes a lightweight, tamper-resistant, and provably secure key agreement protocol designed for remote surgery over the Tactile Internet. The protocol is built upon the computational hardness of the Elliptic Curve Discrete Logarithm Problem (ECDLP), integrates Physical Unclonable Functions (PUFs) to provide hardware-level resistance against cloning and physical tampering attacks, and employs a fuzzy extractor to eliminate PUF output noise, thereby establishing a trusted environment for mutual authentication and secure session key agreement between surgeons and robotic systems. Formal security analysis is conducted using the Real-Or-Random (ROR) model, confirming the protocol’s robustness against standard adversarial models. Performance evaluation demonstrates a 21.48% reduction in computational overhead and a 7.32% reduction in communication cost compared to State-of-the-Art (SOTA) protocols. These results confirm the protocol’s effectiveness for securing resource-constrained, time-sensitive applications in Tactile Internet-assisted remote surgical environments.

Muhammad Asad Saleem +50IEEE Transactions on Mobile Computing
Scholarsynthesis-pnr

SDP: A Unified Protocol and Benchmarking Framework for Reproducible Wi-Fi Sensing

Learning-based Wi-Fi sensing has made rapid progress, yet reproducible evaluation remains difficult because Channel State Information (CSI) measurements are strongly affected by hardware-dependent acquisition factors, preprocessing choices, and heterogeneous subcarrier layouts. This paper proposes the Sensing Data Protocol (SDP), a protocol-level data abstraction and benchmark interface for reproducible OFDM-based Wi-Fi CSI sensing. SDP specifies deterministic signal sanitization, profile-defined baseband frequency-axis projection, canonical tensor construction, and standardized evaluation reporting, mapping heterogeneous CSI measurements into a Canonical CSI Tensor. The protocol is designed to reduce avoidable measurement-side inconsistencies while retaining task-relevant physical variations, with an explicit trade-off between representation stability and information resolution. We evaluate SDP across representative Wi-Fi sensing tasks, including gesture recognition, gait-based identification, fine-grained activity recognition, and activity–location recognition. Controlled analyses compare native-resolution input, principal component analysis (PCA)-based reduction, uniform reduction, and SDP projection under matched splits and seeds. The results show that metadata-driven canonical projection and frequency-dimensionality control both affect sensing performance and computational cost, and that <inline-formula><tex-math notation="LaTeX">$K=30$</tex-math><alternatives><mml:math><mml:mrow><mml:mi>K</mml:mi><mml:mo>=</mml:mo><mml:mn>30</mml:mn></mml:mrow></mml:math><inline-graphic xlink:href="cui-ieq1-3723025.gif"/></alternatives></inline-formula> provides a practical performance–cost trade-off in the evaluated Wi-Fi settings. Additional analyses on phase preservation, representation-level diagnostics, backbone compatibility, and calibrated residential stream inference further show that SDP provides a controlled and reproducible input interface rather than a task-specific accuracy optimizer. Overall, SDP provides a reproducible data and evaluation interface for Wi-Fi sensing, while leaving model architecture and task-specific optimization as independent design choices.

Di Zhang +60IEEE Transactions on Mobile Computing
Scholarsynthesis-pnr

Physics-Guided Diffusion Models for Multi-Scale Prediction of Reference Signal Received Power in Wireless Networks

The Reference Signal Received Power (RSRP) is a crucial factor that determines communication performance in mobile networks. Accurately predicting the RSRP can help network operators perceive user experiences and maximize throughput by optimizing wireless resources. However, existing research into RSRP prediction has limitations in accuracy and verisimilitude. Theoretical derivations and existing data-driven methods consider only easily quantifiable Large-Scale (LS) information and struggle to effectively capture the intertwined LS and Small-Scale (SS) signal attenuation characteristics of the wireless channel. Moreover, the lack of guidance from propagation physics leads to weak accuracy, interpretability, and transferability. In this paper, we propose a diffusion-model-based RSRP prediction model, Channel-Diff. It takes network parameters and multi-attribute urban maps as conditioning inputs, from which large-scale propagation modelling and occlusion–shadow modelling provide LS physical knowledge, while multipath propagation modelling and a microenvironment feature extraction network characterise the conditional statistics of SS fading. We design a physics-guided two-stage training scheme with a noise prior guidance mechanism, enabling effective fusion of multi-scale physical knowledge with the diffusion models. Evaluations demonstrate that Channel-Diff exhibits excellent performance on RSRP prediction, achieving at least 25.15%–37.19% improvement in accuracy relative to baseline methods. Additionally, the model also demonstrates outstanding performance in terms of transferability and training efficiency.

Xiao-Qian Qi +40IEEE Transactions on Mobile Computing
Scholarsynthesis-pnr

Self-Evolving Virtual Network Orchestration From Devices to Cloud Over Shared Infrastructure for Dynamic Services

End-to-end virtual computing network instance (VCNI) orchestration over shared device-to-cloud infrastructure is essential for supporting heterogeneous services with limited physical resources. However, existing network slicing, virtualization, and edge computing approaches usually focus on partial resource management or fixed orchestration intervals, and often neglect adaptive coordination across the full VCNI lifecycle. Motivated by this, we consider a dynamic device-to-cloud VCNI orchestration scenario that jointly accounts for heterogeneous service demands, wireless spectrum resources, cloud computing resources, and reconfiguration overhead. We propose a self-evolving dual-timescale VCNI orchestration framework to coordinate long-timescale inter-VCNI resource provisioning and short-timescale intra-VCNI user scheduling. The framework adaptively adjusts the VCNI reconfiguration frequency according to workload variations, thereby balancing resource consumption, reconfiguration overhead, and task failure penalties. To tackle this problem, we formulate the orchestration process as a hierarchical Markov decision process (MDP) and design a self-adaptive soft actor-critic algorithm, referred to as SaSAC. SaSAC adopts dual-timescale policies and shared twin critics to learn coordinated orchestration decisions under dynamic service states. Experimental results show that SaSAC achieves lower system cost and higher task success rate than learning-based and heuristic baselines. Moreover, additional ablation and post-training results further validate the effectiveness of SaSAC under varying system conditions and dynamic service demands.

Xi-Ying Fan +60IEEE Transactions on Mobile Computing
Scholarsynthesis-pnr

Crowdsensing From a Distance: A Contract Mechanism Design

Mobile crowdsensing (MCS) is a powerful sensing paradigm typically comprising participatory and opportunistic sensing. While participatory sensing incurs high travel costs due to precise movement requirements, opportunistic sensing often fails to ensure task reliability. To address these limitations, we introduce “Crowdsensing from a Distance”, a distance-adaptive sensing scheme that enables users to conduct high-quality sensing from optimized distances relative to task locations. The core of our approach is an optimal contract-based incentive mechanism that integrates sensing distance as a strategic decision variable to balance the non-trivial trade-off between data estimation quality and mobility overhead. To provide a rigorous physical foundation, we derive a multi-user data estimation model that characterizes the analytical relationship between estimation variance and sensing distances, compatible with diverse spatial correlation functions. Furthermore, we address the challenge of asymmetric spatial information by designing a contract that strictly satisfies individual rationality (IR) and incentive compatibility (IC) constraints. This ensures truthfulness in users’ self-selection of contract items matching their actual initial distances without requiring precise real-time coordinates. Our framework is notably model-agnostic, guaranteeing seamless applicability across diverse sensing tasks with heterogeneous cost and correlation functions. The simulations on three real-world datasets and a campus experiment demonstrate that the proposed mechanism significantly outperforms traditional approaches and maintains robust utility growth.

Dong-Ming Luan +20IEEE Transactions on Mobile Computing
Scholarsynthesis-pnr

SRLoc: WiFi Fingerprint Localization Method Against Adversarial Attacks Based on Diffusion-Inspired Model

WiFi fingerprint-based localization using Received Signal Strength Indicator (RSSI) offers deployment convenience and cost efficiency. However, deep learning-based localization models remain highly vulnerable to adversarial attacks, which severely degrade their performance. To address this, we propose SRLoc, a novel diffusion-inspired defense method designed to reconstruct clean RSSI signals from adversarially perturbed inputs. Unlike standard diffusion models relying on multi-step stochastic iterations, SRLoc utilizes a highly efficient single-step conditional denoising network, explicitly using attack intensity as a conditioning vector rather than traditional time steps. Additionally, our architecture replaces the conventional U-Net upsampling module with a 1D transposed convolution tailored for sparse RSSI data. It is opt mized using a customized triple-loss function that imposes physical space and geographical consistency constraints alongside the Huber loss, ensuring physically plausible signal reconstruction. Extensive evaluations on the CJU and Tampere datasets demonstrate that SRLoc successfully defends against adversarial attacks, outperforming four state-of-the-art and five baseline methods. For instance, under a 0.2 attack intensity on the CJU dataset, SRLoc resists 86.5%, 80.8%, and 76.9% of FGSM, PGD, and MIM attacks, exceeding the best competitor by up to 15.8%. On the Tampere dataset, it achieves resistance rates up to 95.7%, surpassing competitors by up to 10.6%. Furthermore, SRLoc maintains excellent defensive capabilities against black-box and unknown attacks, effectively mitigating localization errors and substantially enhancing overall robustness.

Shi-Xun Wu +60IEEE Transactions on Mobile Computing
Scholarsynthesis-pnr

Robust Full-Space Security for UAV-Mounted Multifunctional RIS-Aided Networks With Uncertain Eavesdropper Location

This article investigates robust physical-layer security (PLS) for a downlink wireless network aided by an unmanned aerial vehicle (UAV)-mounted multifunctional reconfigurable intelligent surface (MF-RIS). The location of a passive eavesdropper is modeled by a finite set of candidate positions with known occurrence probabilities, and only statistical channel state information (CSI) of the eavesdropper is available at the transmitter. Different from the existing reconfigurable intelligent surface (RIS) or simultaneous transmission and reflection RIS (STAR-RIS)-aided PLS schemes, the proposed framework jointly exploits the 3-D mobility of the UAV and the amplification/jamming capability of the MF-RIS under this probabilistic uncertainty model. Specifically, a robust weighted sum-rate maximization problem is formulated, where the weighted sum of the legitimate users’ average secrecy and achievable rates is maximized by jointly optimizing the transmit beamforming, the discrete MF-RIS configuration, and the UAV trajectory under power, hardware, and mobility constraints. An efficient alternating optimization (AO) algorithm is developed by combining a weighted minimum mean-square error (WMMSE)-based active beamforming design, a discrete MF-RIS configuration based on cross-entropy optimization (CEO), and a trajectory refinement based on successive convex approximations (SCAs). The simulation results verify the stable convergence behavior of the proposed algorithm and demonstrate consistent weighted sum-rate gains over an equivalent ground-based STAR-RIS benchmark under the same uncertainty model, a UAV-mounted conventional passive RIS baseline, zero-forcing (ZF) and regularized ZF-based baselines, and a WMMSE with semidefinite relaxation (SDR) scheme under different transmit-power, antenna-number, and MF-RIS-size settings.

Ze-Cheng Wang +40IEEE Internet of Things Journal
Scholarsynthesis-pnr

Space–Time–Frequency Multiplexed Directional Transceiver-Based MAC Protocol for Underwater Acoustic Networks

This study proposes an underwater acoustic MAC protocol based on space–time–frequency multiplexing (STFM) to address the tradeoffs among throughput, end-to-end delay, and energy consumption. By integrating time-division access, subband allocation, and multibeam directional communication, the STFM-MAC protocol enables efficient multidimensional resource reuse and enhances parallel transmission without excessive overhead. Compared with existing underwater MAC protocols, STFM-MAC adopts cross-domain cooperative scheduling and parallel link management. The effects of beam configuration, frequency selection, and time-slot allocation on collision probability are analyzed, and a joint optimization framework is designed to balance throughput and latency under long propagation delays. A prototype underwater acoustic modem platform was employed to conduct semi-physical simulations and real-world lake experiments. The experimental results show that STFM-MAC achieves higher throughput and lower end-to-end delay than S-FAMA, MACA-U, and DRLCP-MAC. The maximum throughput increased by approximately 3.02 times, 3.70 times, and 12.85 times, respectively, while the end-to-end delay was reduced by 62.51%, 76.05%, and 92.94%. The consistency between the simulation and experimental results demonstrates the robustness and practical feasibility of the proposed STFM-MAC protocol for underwater acoustic networks (UANs).

Jian-Jun Zhang +40IEEE Internet of Things Journal
Scholarsynthesis-pnr

Coordinated Defense Framework for Mitigating Transmission Line Overloads Induced by VPP-Based Manipulation of Demand Attacks

As an increasing number of distributed energy resources (DERs) are managed by virtual power plants (VPPs), intelligent adversaries could generate abnormal loads through compromised VPPs, inducing line overloads. Such attacks are referred to as VPP-based manipulation of demand (VMAD) attacks. Unlike existing studies that focus mainly on postattack consequence mitigation, this article proposes a coordinated cyber–physical defense framework for mitigating overloads induced by VMAD attacks. First, the features of VMAD attacks in the cyber domain and their resulting overloads in power systems are analyzed. Then, a cybersecurity neighborhood-watch mechanism for VPPs is proposed, which helps dynamically quantify abnormal loads based on the security status of VPPs. Finally, a coordinated dispatch strategy is designed to enhance system resilience to such VMAD attacks, ensuring power flows on transmission lines remain within secure limits. Simulations on several IEEE test systems and a real-world system validate that the proposed coordinated framework can mitigate the overload risk of high-risk lines with lower overall costs, indicating its potential to support the defense against VMAD attacks.

Kang Yan +10IEEE Internet of Things Journal
Scholarsynthesis-pnr

Collaborative PUF and TRNG for SSL-Based IoT Authentication and Key Distribution

A fundamental security challenge in the Internet of Things (IoT) is securing key storage on devices that are deployed in unsecured environments. This challenge which is essential for authentication and key distribution can be effectively addressed by employing physical unclonable functions (PUFs) for key management. We present a collaborative design of PUF and true random number generator (TRNG), both based on semiconductor superlattice (SSL), to enhance the security of IoT key generation, mutual authentication, and key distribution. We established an experimental platform to evaluate the security features of the SSL-PUF and SSL-TRNG. Experimental results show that the SSL-TRNG passes the National Institute of Standards and Technology (NIST) SP 800-22 randomness tests, while the SSL-PUF demonstrates excellent uniformity, uniqueness, twin-similarity, and reliability. The proposed protocol resists multiple attacks, including insider, dictionary, replay, man-in-the-middle (MiTM), quantum computing, machine learning, key compromise, and physical and cloning attacks, while demonstrating notable performance improvements. Specifically, it reduces communication and storage overhead by more than 34.5% and 45.5%, respectively, with low computational complexity and excellent scalability.

Qian Zhang +50IEEE Internet of Things Journal
Scholaraccelerators

Runtime Precision-Adaptive MAC Array Architecture for Energy-Efficient Edge AI Accelerators

Multiply-accumulate (MAC) operations dominate deep neural network (DNN) inference, so the MAC array's energy and latency largely set edge-accelerator efficiency. Conventional fixed-precision arrays waste switching energy computing every product at full bit width regardless of operand size, and waste cycles on every zero-valued activation. This work presents a MAC array with two independently controllable mechanisms addressing both issues: an operand-gated asymmetric-precision MAC that decomposes each signed product into nibble-level partial products and isolates the sub-multipliers a small operand renders redundant, lowering dynamic power; and a per-processing-element zero-skip scheduler that advances each element across only the non-zero taps of its convolution window, reducing cycle count on sparse layers at bit-identical output. The architecture is evaluated on quantized 8-bit MNIST and CIFAR-10 convo-lution layers, implemented on FPGA for resource, timing, and throughput characterization, and synthesized in a 45 nm flow for switching-activity-driven dynamic-power measurement, with independent ablation parameters isolating each mechanism's contribution on otherwise identical hardware. Operand gating removes up to 45.0% of partial-product work on MNIST and up to 31.0% on CIFAR-10, yielding a measured dynamic-power reduction of approximately 11% that raises energy efficiency by 13% (from 204 to 231 GMAC/W) at comparable cell area and a 2.6% frequency cost; the zero-skip scheduler reduces convolution cycle count by 40.3% on sparse layers, where the activation zero fraction reaches 80.2%, at bit-exact output; and the quantization-induced output error stays below 2.5%. The operation-count reduction is data-dependent and traced to per-layer operand statistics, while the dynamic-power benefit is obtained from measured switching activity rather than inferred from utilization.

Pranav Sharma N +10
Scholarmodel-compression

Co-JDPRec: Collaborative Joint Dynamic Pruning for Efficient LLM-based Generative Recommendation

Generative recommendation leverages Large Language Models (LLMs) to model item semantics and user preference dynamics through sequence-to-token generation, but deep Transformer layers and wide Feed Forward Network (FFN) dimensions introduce substantial memory and latency overhead in online recommendation scenarios. Existing LLM pruning methods are mainly designed to preserve general language modeling ability, often overlooking the recommendation-specific states shaped by item semantics, collaborative signals, and temporal preference dynamics. To address this limitation, we propose Co-JDPRec, a collaborative joint dynamic pruning framework that shifts the pruning objective from parameter removal to recommendation-state preservation. Specifically, Co-JDPRec introduces a task-aware saliency criterion that combines collaborative alignment and sequence-order sensitivity, enabling the model to identify neurons critical for user-item relations and evolving preference dynamics. Guided by this saliency, Co-JDPRec jointly searches FFN width and layer depth to derive a compact heterogeneous architecture, followed by distillation recovery to restore recommendation performance. Theoretically, Co-JDPRec characterizes structural importance via first-order Taylor sensitivity under recommendation-aware objectives. Empirically, it achieves superior accuracy–efficiency trade-offs on both title-based and semantic-ID-based generative recommendation backbones, validating the effectiveness of recommendation-state-aware pruning.

Shun-Hang Zheng +50
Scholarmodel-compression

A Systematic Review of Bias, Validation, and Explainability in Artificial Intelligence Algorithms for Dental Diagnostics.

OBJECTIVES To critically evaluate the current role of artificial intelligence (AI) in dental diagnosis across multiple dental disciplines, assessing diagnostic performance, methodological quality, risk of bias, validation, explainability, and clinical applicability. DATA AND SOURCES This systematic review was conducted in accordance with PRISMA guidelines and registered in PROSPERO (CRD42025639340). PubMed/MEDLINE, Scopus, Embase, LILACS, Cochrane Central Register of Controlled Trials, Web of Science, and ClinicalTrials.gov were searched up to December-2025. Two reviewers independently performed study selection, data extraction, and risk-of-bias assessment. Risk of bias in diagnostic accuracy studies was assessed using QUADAS-2-tool, evaluating the domains of patient selection, index test, reference standard, flow and timing whilst randomised studies were assessed using Cochrane RoB 2-tool. Reporting quality and transparency of AI-model development and validation were assessed using the selected items from TRIPOD-AI reporting guideline. Methodological quality and risk of bias in AI model development and evaluation were additionally assessed using PROBAST+AI. STUDY SELECTION Ninety-six studies met the inclusion criteria, of which 90.6% were retrospective. Convolutional neural networks (CNN) were the most frequently used architecture (82.3%). Sixteen studies failed to specify the CNN architecture used. Reported diagnostic accuracy ranged from 81% to 99% across dental specialties, with high performance in implantology (92-99%), orthodontics (88-99%), endodontics (87-99%). Panoramic radiographs and cone-beam computed tomography were the most used imaging modalities. QUADAS-2 analysis indicated that high-risk of bias in 55.2% of studies and most lacked external validation, whilst Assessment using selected TRIPOD+AI items identified important gaps in reporting completeness and transparency, particularly in sample-size justification, missing-data handling, protocol availability, study registration, and data or code availability. CONCLUSIONS AI demonstrates substantial technical capability to enhance diagnostic accuracy and efficiency. However, the predominance of retrospective designs, limited external validation, and methodological weaknesses constrain the strength of current evidence. CLINICAL SIGNIFICANCE The use of AI is potentially considered as a decision-support tool rather than replacement for clinical judgement.

A. Alrayes +20E -journal of dentistry
Scholarmodel-compression

Tree-Compatible Personalized Ranking via Linear Latent-Trait Decomposition

Learning-to-rank systems often use a global scorer despite systematic variation in user preferences. We propose a model-agnostic latent-trait framework for probabilistic ranking objectives. A low-dimensional profile activates shared ranking directions linearly, enabling tractable profile inference with nonlinear document-feature effects. Our iterative route encodes all traits as G profile features in one ranker rather than learning or serving G factor models. Ranker fitting reuses an unmodified standard LTR package, and serving uses one model. An offline module probes the ranker to construct an additive surrogate for profile updates across neural and tree rankers. To our knowledge, this is the first continuous user-specific latent personalization under a ranking loss for standard LambdaMART without architecture changes. In simulation, the method outperforms non-personalized and history-encoder baselines, incurs a modest penalty when G is overestimated, and yields probe-surrogate updates that transfer to the trained ranker. In a retrospective offline Message Search evaluation, personalized LambdaMART outperforms its non-personalized baseline. TreeSHAP provides model-relative trait interpretations.

Pavel Metrikov +40
Scholarmodel-compression

NAS-in-the-Loop: Trajectory-Driven Neural Architecture Search for Safe Autonomous CPS

Neural Architecture Search (NAS) is commonly formulated as a supervised learning optimization problem where architectures are ranked using prediction accuracy. However, in cyber-physical systems, perception models operate with closedloop controllers, and small estimation errors can propagate into unsafe system behaviors. This paper therefore proposes a closedloop trajectory-level NAS framework for autonomous systems using the F1TENTH platform. Instead of optimizing prediction accuracy, the proposed method calls perception modules directly in simulation and calculates cross-track error as the NAS objective. Experimental results on multiple racing tracks show the trajectory-driven NAS selects modules that reduce normalized composite safety-cost relative to accuracy-driven NAS and handcrafted baselines. The largest individual improvement is a 36.8% reduction in cross-track error relative to the handcrafted baseline mean, representing an additional 13.2% reduction compared to the accuracy-driven NAS. The results support the central premise that neural architectures for autonomous systems should be selected using closed-loop behavioral evaluation rather than standalone prediction accuracy.

Zayah Cortright +30Forum on Specification and Design Languages
Scholarmodel-compression

RankNAS: A Pairwise Ranking Model to Accelerate Black-Box Neural Architecture Search

Black-box Neural Architecture Search (NAS) is a robust, model-agnostic technique for creating application-specific neural network architectures. Regrettably, it is often prohibitively expensive as its many candidate architectures must be fully trained to be evaluated. To eliminate this overhead, we propose RankNAS, a lightweight pairwise ranking model that integrates into an evolutionary NAS framework at the candidate evaluation stage, enabling more efficient selection decisions independent of specific search or selection strategies. Contrary to other work, RankNAS encodes neural architectures using a handcrafted, fixed-length embedding rather than a graph encoding, resulting in a significantly simpler encoding that enables faster evaluation. We evaluate RankNAS on the CIFAR-100 dataset in both hardware-constrained and unconstrained settings under fixed low compute budgets of 3 and 10 GPU-hours. Predictor quality is first assessed in isolation: across 10,000 unseen pairwise architecture comparisons, RankNAS achieves ∼83% accuracy with an RMSE of 1.036 fitness units (relative RMSE 7.56%), with most errors occurring when comparing architectures achieving similar performance. The practical impact is then measured by integrating RankNAS into the NAS loop, thereby improving search efficiency and Pareto-front coverage. Overall, the results indicate that RankNAS provides both a reliable ranking signal and a low-overhead mechanism for accelerating black-box NAS under limited compute and hardware constraints.

Vasilis Moschou +20Coins
Scholarmodel-compression

AI-Assisted Sonography for the Diagnosis of Thyroid Nodules.

PURPOSE To evaluate the diagnostic accuracy, clinical utility, and workflow integration of artificial intelligence (AI)-assisted sonography for the assessment of thyroid nodules and to assess the tool's potential for improving diagnostic consistency, reducing the number of unnecessary biopsies, and assisting with clinical decision-making. METHODS Database searches were conducted to identify studies published from January 2018 to June 2025. Eligible studies compared AI-based models for thyroid sonography using histopathology or fine-needle aspiration (FNA) as reference standards. Data extraction was conducted according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses 2020 criteria. Variables included AI architecture, sonography mode or technique, study design, clinical impact measures (ie, FNA reduction, time savings), and diagnostic metrics (ie, sensitivity, specificity, and area under curve). RESULTS A total of 30 studies were included. Convolutional neural networks were the most common architecture. Area under curve values ranged from 0.78 to 0.97, specificity from 70.4% to 90.7% and sensitivity from 75.6% to 94.0%. Clinical advantages included a 10% to 45% reduction in the number of FNAs performed, with some studies reporting between 2.5 to 3.1 minutes of interpretation time saved per case. Risk-of-bias assessment determined 53.3% of the studies to be low risk, 33.3% to be moderate risk, and 13.3% to be high risk, primarily because of the retrospective design or small sample size. DISCUSSION Thyroid nodules are prevalent; however, interpretation of sonograms is prone to interobserver variability. Results from this systematic review indicate that AI-aided sonography exhibits high levels of diagnostic efficacy and quantifiable clinical advantages, especially when combined with the American College of Radiology Thyroid Imaging and Reporting Data System and externally validated. CONCLUSION This review maps the evolution of experimental models to clinically implementable systems and calls for prospective multicenter trials, standardized reporting, and explainable AI to facilitate safe and reproducible usage in routine thyroid imaging. These findings support Sustainable Development Goal (SDG) 3 (good health and well-being) by promoting higher-quality diagnostic care and SDG 9 (industry, innovation, and infrastructure) by advancing AI in medical imaging.

Javeria Khan0Radiologic Technology
Scholarmodel-compression

A Comparative Study of Intelligent Methods for Steel Surface Defect Classification in Industrial Quality Control

The intelligent transformation of manufacturing processes increasingly relies on automated visual and parametric inspection of intermediate products, where machine learning models classify surface defects from sensor-derived features. Despite the rapid progress of deep tabular learning architectures, the relative performance of classical ensemble methods, gradient boosting, and modern neural-attention models on small industrial datasets remains insufficiently characterized. The absence of unified comparison protocols hinders informed deployment decisions in industrial quality-control environments. This paper presents a comparative study of five representative intelligent methods — logistic regression, Random Forest, CatBoost, a multilayer perceptron, and TabNet — applied to the classification of steel-plate surface defects across seven fault categories. A unified evaluation protocol is proposed, integrating correlation-based feature selection, synthetic minority oversampling within cross-validation folds, exhaustive grid-search hyperparameter tuning under stratified cross-validation, and held-out test-set evaluation. The protocol prevents data leakage at every stage and treats class imbalance through SMOTENC applied inside the resampling loop. The experimental results show that ensemble methods — Random Forest and CatBoost — achieve closely matched leading performance (test F1-macro 0.81) The multilayer perceptron and TabNet trail by 4-8 F1-macro points. Per-class analysis identifies systematic confusion among visually similar defect categories. Feature-importance analysis attributes the predictive signal primarily to defect-size descriptors and production-line geometry. The proposed protocol provides a reproducible framework for selecting intelligent methods in resource-constrained industrial quality-control deployments.

N. Grechanikov +20
Scholaroptimization

DeltaUI: Framework-Normalized UI State Transition Modeling for Multi-Task Front-End Engineering

Multimodal Large Language Models (MLLMs) are emerging as a foundation for automated front-end engineering because they enable joint reasoning over screenshots and natural-language instructions. However, most front-end assistants treat this workflow as a set of isolated code-generation tasks, directly decoding framework-specific syntax for each request. This direct-decoding paradigm leaves three limitations unresolved: (i) it provides no transferable prediction target across tasks and frameworks, (ii) it offers no stable representation for localized edits and repairs, and (iii) it provides no auditable path from multimodal input to executable outcomes. In this paper, we propose DeltaUI, a unified framework that recasts generation, editing, and repair as predictions of transitions in a shared, framework-normalized UI state space, modeling UI change rather than directly decoding target code. DeltaUI begins by parsing heterogeneous front-end artifacts into a typed UI state graph that separates shared UI semantics from framework-specific surface syntax. It then uses a multimodal transition model to predict structured state operations by encoding the source state and multimodal evidence (e.g., screenshots, instructions, and issues), preserving explicit source-to-update correspondence. A deterministic framework-conditioned realizer converts the resulting state into target-framework code. Finally, an execution-aware critic drives a bounded refinement loop using compiler and render feedback, ensuring the full path remains auditable. Extensive experiments on DesignBench and Design2Code demonstrate that modeling UI-state transitions yields consistent gains over direct code-decoding baselines across multiple MLLM backbones while retaining effective transfer to real-world webpages. The code and implementation of DeltaUI are provided in the supplementary material.

Jia-Ye Lin +80
Scholaroptimization

MemDTG: Dynamic Temporal Memory Graphs for Evidence-Admissible Agent Memory

Long-term agent memory is an online, non-monotonic retrieval problem. New interactions may reinforce, contradict, supersede, specialize, abstract, or expire earlier memories; the same fact can be a wrong current answer, a correct historical answer, or necessary transition evidence depending on query intent. Existing chunk, summary, and graph memories retrieve by semantic or graph relevance and treat time as post-hoc metadata, so they cannot determine whether a memory is observable, valid, lifecycle-admissible, and provenance-supported before it enters the candidate space. We propose MemDTG, a dynamic temporal memory graph that frames agent memory as temporal evidence admission. MemDTG separates immutable raw episodes from a mutable entity-state temporal graph and a memory-evolution graph over assertion versions. State slots localize mutable user and task states; an LLM-assisted slot-local resolver classifies semantic evolution only within bounded temporal histories; and deterministic commit rules update valid intervals, lifecycle states, and typed evolution edges. At read time, a query compiler instantiates observed-time, valid-time, version, edge-type, and role constraints before ranking, and QV-CTW retrieves evidence paths only inside the query-legal temporal subgraph. Evidence certificates record selected facts, sources, roles, and rejected stale or future alternatives, making retrieval auditable. On LongMemEval and LoCoMo, MemDTG achieves the best LLM-judge accuracy and Citation F1. Ablation and temporal error analyses confirm that provenance certificates and legal-subgraph pre-filtering are each essential for these gains, reducing stale-answer errors by 61.7% and future leaks by 80.1%.

Ya-Xin Deng +40
Scholaroptimization

GraphCraft: Skill-Centric Curation of Evolving GraphRAG Pipelines

GraphRAG pipelines build entity–relation graphs from text and use them for retrieval and question answering, but extraction noise, entity fragmentation, and unstable clustering can silently corrupt the evidence behind an answer. Correcting these errors is difficult to scale because their dominant forms depend on the corpus and repeated corrections remain manual. We present GraphCraft, an interactive system that records provenance-rich human corrections and compiles repeated corrections into reusable curation skills. Each skill identifies similar errors in subsequent corpus updates and proposes edits for curator review; proposals are previewed rather than applied automatically. The same journal links corrections and skill suggestions back to their graph context and source spans, while bi-temporal graph state distinguishes changes to the stored graph from changes in the underlying world. We evaluate a compiled skill on held-out ingestion batches and demonstrate the complete workflow from diagnosing an erroneous answer, through repairing its graph evidence, to reusing the correction on later data. GraphCraft is released as an open platform with offline deployment. Code and offline deployment: https://github.com/apsolutions/graphrag; demo video: https://vimeo.com/1141894235.

Ilia Karpov +50
Scholaredge-ai

A Challenge–Response Remapping Reliability Enhancement Method for APUF Based on Adjacent Challenge-Bit Flipping

With the rapid development of the Internet of Things (IoT), security issues in resource-constrained devices have become increasingly critical. Physical unclonable function (PUF) exploits unavoidable process variations introduced during IC manufacturing to generate unique responses and has been widely used for device authentication and cryptographic key generation. Among various PUF architectures, the arbiter PUF (APUF) can generate a large number of challenge–response pairs (CRPs) with low hardware overhead. However, its reliability is highly sensitive to environmental variations and device aging, which may degrade the security of authentication protocols. In this article, a challenge–response remapping reliability enhancement method for APUF based on adjacent challenge-bit flipping is proposed. First, a machine learning (ML) algorithm is employed to construct a software model of the APUF. A delay difference threshold is then introduced to identify challenges whose accumulated propagation delay differences fall below the threshold. Subsequently, specific adjacent bits in these challenges are flipped to transform them into highly reliable challenges. The proposed method is evaluated through Python-based simulations and FPGA-based hardware measurements. Experimental results show that the simulated reliability of the APUF can be improved to 100% under three environmental noise conditions. Moreover, within a temperature range from $- 20~^{\circ }$ C to $80~^{\circ }$ C, the measured reliability of the APUF can also reach 100% without reducing the number of available CRPs. Therefore, the proposed method significantly improves the reliability of APUFs without introducing additional hardware overhead or reducing the number of available CRPs, providing an effective solution for high-reliability authentication in IoT devices.

Yan-Kun Lin +50IEEE Internet of Things Journal
Scholaredge-ai

Hierarchical Feature Fusion and Cross-Scale Coding for Machine Vision Tasks in Internet of Things

In Internet of Things (IoT) applications, image compression for machines (ICMs) aims to reduce the amount of data transmitted for machine vision tasks such as object detection and instance segmentation. Recently, the compression of intermediate features has attracted increasing attention in the ICM because it enables compact representations of task-relevant information. However, existing methods typically fuse multiscale features into a single compact representation, which limits the representational capacity and results in suboptimal rate-accuracy (RA) performance. To address this, we propose an end-to-end cross-scale feature compression (CSFC) framework for compressing intermediate features within the ICM. The proposed CSFC framework consists of a hierarchical feature fusion (HFF) network, a cross-scale entropy coding (CSEC) network, and a hierarchical feature reconstruction (HFR) network. Specifically, our HFF network first fuses multiscale features into two compact representations. Then, the CSEC network reduces redundancy via cross-scale conditional coding to improve coding efficiency. Finally, the HFR network reconstructs multiscale features for machine vision tasks. Experiments on object detection and instance segmentation demonstrate that our method achieves superior RA performance, yielding Bjøntegaard delta rate (BD-rate) reductions of 99.90% and 99.89% over the VVC-based feature anchor, and 78.11% and 71.65% over the learning-based L-MSFC anchor, respectively.

Lei Xiong +60IEEE Internet of Things Journal
Scholaredge-ai

Verifiable Quantum Machine Learning Inference Consistency for Classical Clients in Noisy Quantum Clouds

Quantum machine learning (QML) is an emerging paradigm that aims to achieve high expressive power by leveraging quantum advantage. However, realizing these advantages in practice inevitably requires executing inference on remote, noisy quantum hardware. Such hardware produces probabilistic and nonreproducible outcomes, leaving classical users without a principled way to verify reported results. We propose zkQML, the first framework that enables a fully classical verifier to certify that a reported QML inference result is consistent with a committed model up to a device-calibrated noise tolerance, by leveraging zero-knowledge succinct noninteractive arguments of knowledge (zk-SNARKs). zkQML arithmetizes the ideal semantics of quantum inference and enforces consistency using cryptographic zk-SNARKs, while explicitly accounting for device noise through a noise-bounded correctness notion. Experiments on both simulators and real quantum hardware show that the proposed arithmetic modeling closely matches ideal quantum inference and enables verifier-efficient verification with small proof sizes and near-constant verification time. These results demonstrate that zkQML provides a practical and sound foundation for trustworthy quantum inference in realistic, nonreproducible execution environments.

Seung-Kwan Lee +40IEEE Internet of Things Journal
Scholaredge-ai

UnlearnKT: Machine Unlearning for Privacy-Preserving Knowledge Tracing in E-Learning IoT Systems

Knowledge tracing (KT) is a core component of E-Learning Internet of Things (E-IoT) systems, where large-scale learner-question interaction sequences are continuously collected to estimate knowledge states and support personalized educational services. However, KT models may retain information about specific learners even after their data has been removed, creating privacy and compliance risks. To address this problem, we first present UnlearnKT, a KT-specific machine unlearning framework that organizes learner-level deletion requests into standard training, targeted unlearning, and postunlearning evaluation. Within this framework, we propose Question-Exposure-Aware Fisher Unlearning for Privacy-Preserving KT (QEFU-KT), a KT-specific retraining-free method that estimates question-level exposure contrast between the forget and retain partitions, strengthens Fisher perturbation on forget-heavy interaction embeddings and prediction rows, and then applies bounded targeted ascent to the most exposed question parameters. Extensive experiments on six public KT datasets and three representative KT backbones show that QEFU-KT remains in the strongest adaptive unlearning score (AUS) tier (mean AUS = 0.9242) while achieving an average RunTime per 1k forgotten interaction (RT/1k-Int) of only 15.077, which places it in the fastest cost band and well below retraining-heavy baselines such as GoldModel (100.512), selective synaptic dampening (SSD) (74.470), and Finetuning (74.469). These results show that QEFU-KT offers the strongest overall quality-efficiency tradeoff for privacy-preserving KT in E-IoT systems.

Tao He +50IEEE Internet of Things Journal
Scholaredge-ai

FedTQK: Communication-Efficient Quantum Kernel Federated Learning for Anomaly Detection in Resource-Constrained IoT Networks

We propose Federated Trainable Quantum Kernel (FedTQK), a federated quantum kernel learning (FQKL) framework for anomaly detection in resource-constrained Internet of Things (IoT) networks. FedTQK is an ultralightweight hybrid detector: a local classical autoencoder (AE) compresses raw traffic, only 24 parameters of a shared parameterized quantum circuit (PQC) are exchanged with the server each round, and anomalies are decided by a one-class support vector machine (OCSVM) defined on the resulting quantum kernel. The distinguishing property of FedTQK is not detection accuracy but the per-round communication footprint. Each per-round upload occupies 96 uplink bytes per client, which fits inside a single low-power wide-area network (LPWAN) uplink frame at moderate spreading factors (SFs); measured across a full 20-round, ten-client training schedule, the aggregate uplink drops from 1536.7 KiB (a classical federated averaging (FedAvg) AE baseline evaluated on the same schedule) to 18.75 KiB. This reduction is a structural property of quantum kernel embeddings, whose parameter count grows with the number of trainable rotations rather than with the size of a classical weight tensor, and it is therefore expected to extrapolate to substantially higher dimensional detection problems in which transmitting a classical model each round would exceed the IoT device duty cycle. FedTQK is intended for the regime in which the communication budget is the binding constraint and accuracy parity with strong classical baselines—not accuracy dominance—is the acceptable ceiling. Experiments on NSL-KDD confirm that FedTQK attains an accuracy that numerically overlaps with a classical FedAvg AE baseline and with a purely local, nonfederated reference. The results quantify the communication reduction relative to the classical baseline and identify the local nonfederated variant as the natural reference against which the value of federation should be judged.

Sungkwan Youm +20IEEE Internet of Things Journal
Scholaredge-ai

A Textile-Based Wilkinson-Divider RF Sensor for Finger-State Extraction in Wearable IoT Systems

Gesture-aware wearable interfaces for human-centric Internet of Things (IoT) systems require compact sensing front ends that can provide finger-level observability without dense wiring, bulky sensor arrays, high-dimensional RF measurements, or computation-heavy learning models. This work presents a textile guided-wave RF sensing method that repurposes a glove-integrated <inline-formula> <tex-math notation="LaTeX">$1\times 4$ </tex-math></inline-formula> Wilkinson power divider as a four-channel finger-posture transducer at <inline-formula> <tex-math notation="LaTeX">$865{\,}\mathrm {MHz}$ </tex-math></inline-formula>. Unlike its conventional role as a passive RF power-distribution network, each output branch is routed along an individual finger region so that extension/flexion perturbs the local guided-wave propagation path and modulates the corresponding input-to-branch transmission magnitude. The structure, fabricated on a flexible cotton substrate, provides four synchronized branch-magnitude observables, <inline-formula> <tex-math notation="LaTeX">$\lvert S_{21}\rvert $ </tex-math></inline-formula>, <inline-formula> <tex-math notation="LaTeX">$\lvert S_{31}\rvert $ </tex-math></inline-formula>, <inline-formula> <tex-math notation="LaTeX">$\lvert S_{41}\rvert $ </tex-math></inline-formula>, and <inline-formula> <tex-math notation="LaTeX">$\lvert S_{51}\rvert $ </tex-math></inline-formula>, associated with Ports 2–5. The branch responses are acquired using software-defined radios (SDRs) and processed through a deterministic edge-processing pipeline comprising signal conditioning, denoising, and adaptive-hysteresis digitization to generate compact binary finger states. An application-level lookup rule then maps the resulting 4-bit state vector to commands for live dashboard visualization. Measurements confirm good multiport matching, with all port reflection coefficients below <inline-formula> <tex-math notation="LaTeX">$-10{\,}\mathrm {dB}$ </tex-math></inline-formula> at <inline-formula> <tex-math notation="LaTeX">$865{\,}\mathrm {MHz}$ </tex-math></inline-formula>, near-balanced input-to-branch transmission responses, and repeatable posture-dependent magnitude modulation. The extracted branch responses are subsequently converted into representative per-finger binary states and streamed as compact state packets for real-time monitoring. The results support the divider-as-sensor principle as a compact proof-of-concept route toward textile-integrated IoT wearables for gesture-aware human–machine interaction (HMI), human–robot interaction (HRI), and connected control applications.

Amir Sohail +50IEEE Internet of Things Journal
Scholaredge-ai

Semantically Guided Cross-Subject EEG Emotion Recognition with Pretrained Language Model

EEG emotion recognition typically follows the standard signal-to-label paradigm. It maps EEG signals to discrete labels via feature engineering and machine learning, yet neglects the rich semantic information inherent in the task. This purely data-driven approach easily overfits to subject-specific physiological patterns, severely degrading its generalization on new subjects. Here, we propose a Semantically Guided Domain Adaptation (SGDA) method for cross-subject EEG emotion recognition. Specifically, SGDA projects non-stationary EEG features into an emotional semantic space, leveraging the discriminative power of structured linguistic embeddings to enhance class separability. To construct this unified space, a pretrained language model is utilized to extract linguistic embeddings from emotion text labels, which serve as invariant semantic prototypes to guide cross-domain feature alignment. Moreover, to facilitate robust inference, an adaptive distribution-aware feature fusion strategy dynamically aggregates multi-source knowledge based on the distances between target samples and source centroids. Extensive experiments on five benchmarks demonstrate that SGDA achieves state-of-the-art performance in cross-subject scenarios. The source code is publicly available at https://github.com/EDH777/SGDA.

Jing-Cong Li +60
Scholaredge-ai

Machine Learning-Based IoT Intrusion Detection: A Survey and Performance Comparison

As IoT devices multiply, the cyber attack surface has widened just as fast, and with it the need for intrusion detection systems (IDS) that are efficient, dependable, and protective of user privacy. This paper organizes IoT IDS methods into a taxonomy built around training architecture and reviews them under two main paradigms, misuse detection and anomaly detection, with the focus on centralized traditional machine learning and deep learning. Experiments on several public datasets point to a clear pattern: in centralized settings, tree-based ensemble models and the hybrid CNN-LSTM architecture give the strongest overall results, whereas Isolation Forest holds up better once the data is compressed at the edge. For federated deployment, tree-based ensembles stand out as the preferred local classifiers.

Rui-Xi He0
Scholaredge-ai

Pooled Spatio-Temporal Machine-Learning Forecasting of Ambient CO2 Using IoT-Based Meteorological Sensor Networks: A SHAP-Based Explainability Study Across Heterogeneous Greek Sites

Forecasting ambient CO2 concentration at the local scale remains comparatively under-studied relative to emissions or global background-concentration forecasting, particularly with respect to a direct comparison of pooled (data from multiple sites combined into a single model) versus site-specific (the same stations, modeled independently) strategies over heterogeneous IoT sensor networks. This work addresses that gap by developing and evaluating a unified machine-learning framework on data from ten agrometeorological stations of the Hellenic National Meteorological Service, geographically distributed across the Greek territory and spanning six environmental typologies (coastal, island, urban/industrial, lowland/rural, semi-urban, and mountain/forest environments), using 5 min data from the spring season (1 April–2 June 2026). Five model families—Persistence, ARIMA/SARIMAX, Random Forest, XGBoost, and LightGBM—were evaluated under a strictly chronological data-splitting scheme (40 sliding train/validation pairs, a held-out 20% test set) and a systematic paired statistical testing procedure (Shapiro–Wilk, paired t-test, Wilcoxon signed-rank) for selecting the best hyperparameter combinations, in both pooled and site-specific configurations. The results show that at eight of the ten stations, the simple persistence baseline prevails, owing to the strong autocorrelation of CO2 concentration in stable environments, whereas at the two mountain stations (Vitsi, Parnetha), where CO2 dynamics are inherently more unstable, machine-learning models outperform it (at Vitsi, RMSE 5.73 versus 7.46 ppm and R2 0.84 versus 0.73). SHAP analysis confirms and explains this pattern, revealing a shift of predictive weight away from the lagged CO2 value and toward relative humidity and temperature precisely in the unstable environments. Furthermore, the pooled strategy proved more robust than the site-specific one for every machine-learning model (e.g., pooled Random Forest RMSE = 5.43 versus a mean of 16.20 for the site-specific models), mainly because it is less affected by seasonal level shifts beyond the training range; the pooled machine-learning models did not, however, outperform persistence. For the spring period studied, the findings indicate that the value of artificial intelligence, and especially of explainable artificial intelligence, is not universal but concentrates precisely where a stable, predictable environment is absent—a finding with direct implications for the targeted design of future CO2 monitoring systems.

Y. Kiouvrekis +30Computers
Scholaredge-ai

Contextual Contrastive Representation Learning with LightGBM for Intrusion Detection in Heterogeneous IoT Networks

The increasing adoption of Internet of Things (IoT) technologies in smart farming has improved agricultural automation and decision-making while expanding the attack surface of interconnected devices and communication networks. Intrusion detection in these environments is challenged by heterogeneous network features, severe class imbalance, and evolving attack patterns. This study proposes CCRL-LightGBM, a contextual contrastive representation learning framework that integrates TabTransformer-based feature learning, supervised contrastive refinement, and Light Gradient Boosting Machine (LightGBM) classification. The framework was evaluated on the CICIoT2023 and CIC IoT-DIAD 2024 datasets using a unified preprocessing pipeline and compared with representative machine learning and deep learning models. Under random-stratified evaluation, CCRL-LightGBM achieved an accuracy of 99.63% and F1-score of 99.81% on CICIoT2023 and an accuracy of 99.39% and F1-score of 99.68% on CIC IoT-DIAD 2024. GPU inference latency was 3.78 ms per sample under the evaluated experimental conditions. In paired comparisons with TT–LightGBM, the CCRL-LightGBM configuration achieved higher F1-scores under temporal, source-disjoint, and device-disjoint evaluation protocols, with device-disjoint F1-scores of 93.31% on CICIoT2023 and 92.56% on CIC IoT-DIAD 2024. However, performance declined under stricter partitioning protocols, highlighting the challenges of generalization to temporally shifted traffic and previously unseen source or device groups. Unadapted cross-dataset evaluation further revealed limitations in transfer performance between the two benchmark environments. These results demonstrate strong benchmark-level performance and improved results relative to the baseline under the evaluated conditions while highlighting the need for further validation using agriculture-specific traffic and deployment-level measurements before operational suitability for smart-farming environments can be established.

Sunil Kumar +70Future Internet
Scholarai-hardware

Efficient Decentralized Modular Continual Learning via Aligned Modularization and Selective Collaboration

With the deep integration of Internet of Things and Artificial Intelligence, End Devices (EDs) must achieve long-term autonomous learning in environments characterized by limited computational and communication resources, continuously evolving tasks, and highly heterogeneous data. Existing single-device Continual Learning (CL) methods are constrained by ED computing, storage, and local data, making it difficult to ensure model scalability and learning efficiency in resource-constrained environments. Meanwhile, distributed CL approaches face new systemic challenges such as asynchronous task arrival, device and data heterogeneity, and high communication overhead. To address these challenges, this paper proposes a Decentralized Modular Continual Learning (DeModCL) framework. It achieves module-level collaborative learning without central coordination by constructing reusable modules across EDs. At the single-device level, DeModCL combines distributed drift detection with historical memory trace modeling to introduce a Memory-Trace-based Dynamic Routing (MTDR). MTDR adaptively reconfigures module architecture and routing upon new task arrival. At the cross-device level, we introduce Module Capability Signature via Task Anchors (MSA) to enable EDs to identify relevant and shareable modules. Building upon this, we design a Collaborative Module Decision (CMD) that balances learning performance gains against communication overhead, achieving efficient and selective module sharing. Finally, experiments are conducted on vision, sensor-based activity recognition, and real-world continuous mobile sensing datasets to evaluate the proposed method under heterogeneous CL workloads. On the system-efficiency benchmark over CIFAR-100, Tiny-ImageNet, CUB-200, and UCI-HAR, DeModCL improves efficiency by 70.53%–377.82% over representative distributed CL baselines, where efficiency is measured by average accuracy over system continual-learning completion time.

Ming-Ze Li +50IEEE Transactions on Mobile Computing
Scholarai-hardware

Scaling LLMs via Collaboration: Efficient LLM Inference Across Resource-Constrained Devices

Large language models (LLMs) deliver remarkable capabilities but remain difficult to deploy on memory- and compute-limited edge devices. A natural question is whether multiple devices can collaboratively execute token-by-token inference with low latency. Existing parallelism schemes, designed for datacenter GPUs and batch-oriented workloads, break down on edge devices due to ill-suited partitioning and heavy synchronization overheads. We identify three core challenges: (i) per-token synchronization dominates latency in decoding, (ii) naive scaling amplifies communication cost, and (iii) Transformer operators harbor hidden heterogeneous computation–communication characteristics that remain underutilized. Our key insight is that partitioning and synchronization must be co-optimized to unlock scalable edge collaboration. We present CoSEI, a system that introduces operator-aware partitioning and synchronization-minimizing execution, guided by a lightweight planner and a tailored runtime. A prototype built atop llama.cpp achieves up to 5.4× lower latency than pipeline baselines on an eight-node edge device cluster, while enabling models that exceed single-device memory. Results demonstrate that with the right design, collaborative edge inference can make LLMs practical beyond the cloud.

Fei Zeng +50IEEE Transactions on Mobile Computing
Scholarai-hardware

AgileAirNeuro: Mobility-Robust Over-the-Air Neural Inference for Edge Intelligence

The growing demand for high-dimensional artificial intelligence (AI) services on resource-constrained Internet of Things (IoT) terminals has driven the development of distributed edge intelligence. Among these enabling technologies, split inference has attracted significant attention due to its capability for lightweight deployment on edge computing devices. Meanwhile, over-the-air computation (AirComp), which exploits the principle of analog waveform superposition, has catalyzed the “over-the-air neural computing” paradigm to achieve end-to-end split inference across communication networks. However, high-mobility distributed inference can be limited by the spatial rank available at a single access point (AP) and by Doppler-induced phase mismatch. Limited spatial rank can truncate high-dimensional features, while Doppler shifts can impair coherent waveform aggregation. To address these challenges, we develop AgileAirNeuro, which combines heterogeneous multi-AP AirComp with sensed-velocity-based phase compensation. Under the stated nonzero-gain, generic-angle, and path-richness assumptions, we derive an ideal rank-additivity result that characterizes the algebraic spatial-DoF upper bound of heterogeneous multi-AP cooperation. Within the local residual-phase regime, we derive a first-order computation-mse model and its Cramér–Rao bound (CRB)-induced lower bound surrogate, relating velocity-estimation uncertainty to neural computation distortion. We formulate a stochastic computation-mse minimization problem and develop a spatio-temporal alternating relaxation (STAR) approach that combines greedy rank-aware topology selection with alternating optimization of the local surrogate. Simulations under the specified satellite–terrestrial integrated network (STIN) configurations show lower computation error than the evaluated baselines. Under preconfigured topology and transceivers, the simulated online air-interface feature-transmission latency is 92.65% lower than digital offloading. In the same simulations, AgileAirNeuro achieves higher inference accuracy than terrestrial baselines in the tested high-dimensional and line-of-sight (LoS)-dominant cases.

Pei-Bo Wang +20IEEE Internet of Things Journal
Scholarai-hardware

Design Method of Endogenous Secure T-BOX Based on Lightweight Dynamic Heterogeneous Redundant Environment

The telematics box (T-BOX) is the core component for connecting intravehicle and intervehicle networks, which is also the component extremely vulnerable to cyberattacks, especially from unknown vulnerabilities and backdoors. Given the limited computing and storage resources of T-BOX, improving its resistance to unknown cyberattacks remains challenging. To solve this problem, an endogenous secure T-BOX architecture based on lightweight dynamic heterogeneous redundant environment (LDHRE) is proposed. First, the lightweight dynamic heterogeneous redundant architecture and operation mechanism of T-BOX are established. Then, the probabilistic model for the contextual information anomalies of executors is given, including confidence coefficient, running environment resources, processes, and files. An arbitration method (CMIA) based on multidimensional contextual information of the executor is proposed, assisting the arbiter to output the final decision quickly and accurately. The experimental results show that the LDHRE T-BOX can effectively resist unknown cyberattacks. Compared with a traditional T-BOX, the LDHRE T-BOX shows a 2.43% higher memory usage and a 2.51% higher CPU usage. When multiple executors are attacked simultaneously, the accuracy of CMIA in the LDHRE T-BOX is higher than that of majority voting under different attack scenarios. In summary, the validity and feasibility of the LDHRE and the CMIA have been verified.

Jin-Feng Zhang +50IEEE Internet of Things Journal
Scholarai-hardware

Memory-Aware Hypergraph Reasoning for Propagation-Aware Trustworthy Content Service in Social Internet of Things

Trustworthy content service in the Social Internet of Things (SIoT) requires joint reasoning over propagation structure, credibility, uncertainty, and service constraints. Existing methods usually treat these functions separately. This article proposes memory-aware hypergraph trust (MHT), which couples next-user prediction in reconstructed conversations, credibility classification, and rule-derived service ranking. MHT represents each event as a masked typed hyperedge and retrieves learnable propagation prototypes that summarize historical interaction patterns. Bounded relation gates and normalized node–hyperedge–node propagation produce a shared state, while an auxiliary transition branch regularizes consecutive service states. On PHEME and CoAID, MHT improves the strongest evaluated baselines by 3.19–5.08 percentage points across five reported metrics. Two-sided Welch tests computed from five-run summary statistics remain significant after Holm correction ( $p_{\mathrm {Holm}}\leq 1.73{\,}\times {\,}10^{-5}$ ). Cached online inference requires 14.2M parameters, 3.5G FLOPs, 18.5 ms per sample at batch size 64, and 2.6 GB peak GPU memory. PHEME and CoAID are social-media benchmarks without physical SIoT telemetry. The evidence, therefore, supports a propagation-aware decision-support formulation, not a deployed SIoT controller.

Zheng-Shun Li +10IEEE Internet of Things Journal
Scholarai-hardware

Evaluating the robustness of explainable AI methods to adversarial attacks in intrusion and malware detection

Intrusion detection systems are increasingly paired with eXplainable AI (XAI) so that security analysts can understand and act on machine learning and deep learning model decisions. The explanations themselves, however, may be fragile: a perturbed input can change not only the predicted label but also the feature-importance narrative that justifies it. This paper presents a systematic evaluation of how two gradient-based untargeted evasion attacks, the fast gradient sign method and projected gradient descent (PGD), affect three widely used explainers in intrusion and malware detection: DALEX for dataset-level global importance, SHAP at both global and local granularity, and Integrated Gradients (IG) for local attribution. We quantify explanation stability with three complementary measures, Top-k feature overlap, cosine similarity between attribution vectors, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L_1$$\end{document} attribution distance, computed over the full test set rather than on individually selected samples. We further separate two effects that are commonly conflated; explanation drift is reported separately for adversarial inputs whose predicted label changes and for those whose predicted label is preserved, which isolates explanation instability from prediction instability. Experiments on CICMalMem22 (obfuscated malware memory forensics) and CICIDS17 (network intrusion) show measurable but method- and dataset-dependent drift when predictions are preserved; the mean Top-5 overlap ranges from 0.586 to 0.892, while cosine similarity ranges from 0.875 to 0.962. The strongest preserved-prediction drift occurs for SHAP on CICMalMem22 under untargeted PGD. These findings indicate that XAI constitutes an additional attack surface in cybersecurity pipelines and that evaluating an IDS should include explanation-robustness measures alongside accuracy and F1-score.

Malik al-Essa +50Discover Artificial Intelligence
Scholarai-hardware

Resource-Efficient Generative AI: Rethinking Knowledge Management and Delivery

Generative AI is increasingly embedded in information retrieval, knowledge management, and recommender systems, yet its adoption is constrained by finite compute, memory, latency, energy, and data-governance budgets. This tutorial presents a resource-aware and sovereign perspective on generative AI for scalable knowledge access, framing efficiency as a multi-objective optimization over accuracy, latency, and cost. We structure the tutorial around: intelligence building, which addresses training and adaptation under explicit resource budgets, and intelligence delivery, which focuses on serving-time inference under strict tail constraints (e.g., P99 latency, peak memory) typical of large-scale retrieval and decision-support pipelines. Core techniques, including parameter-efficient fine-tuning, sparsity and pruning, quantization, distillation, mixture-of-experts routing, and speculative decoding, are analyzed as controlled approximations that trade utility and predictive uncertainty. The tutorial equips participants with methods and deployable design patterns for building cost-efficient, privacy-preserving generative AI systems aligned with CIKM's core themes of scalable, trustworthy knowledge access under real constraints.

João Gama +10
Scholarai-hardware

LuWu: Adaptive Layer-wise Heterogeneous KV Cache Quantization

The KV cache is essential for efficient autoregressive inference in Large Language Models (LLMs), but its memory cost increases with context length, layer depth, and request concurrency, limiting long-context and high-throughput serving, particularly in Computing Power Network (CPN) environments. KV cache quantization reduces this cost by lowering the storage bit-width of keys and values. However, existing methods often rely on uniform precision or static mixed-precision policies, which cannot fully capture layer-wise differences in quantization sensitivity or decoding-time changes in precision requirements. This paper proposes LuWu, an adaptive layer-wise heterogeneous KV cache quantization method. LuWu first performs offline calibration to measure attention-output errors under different K/V precision pairs, clusters layers with similar error response patterns, prunes unsuitable candidate configurations, and searches an initial heterogeneous precision configuration under accuracy and bit-width objectives. During online inference, LuWu uses the Layer Activity Score (LAS) to adjust the precision of selected layers, while a hierarchical residual-code representation supports efficient promotion and demotion between precision levels. Experiments on representative reasoning, long-context, and language modeling tasks show that LuWu achieves a favorable accuracy--bitwidth trade-off. Compared with representative state-of-the-art K8V8 baselines, LuWu reduces the average equivalent bit-width by 29.5%--38.1% while maintaining competitive inference accuracy. These results highlight the effectiveness of adaptive layer-wise heterogeneous quantization for memory-efficient LLM inference.

He-Xuan Lv +20