KiCAD-MCP-Server
KiCAD MCP is a Model Context Protocol (MCP) implementation that enables Large Language Models (LLMs) like Claude to directly interact with KiCAD for printed circuit board design.
9 repos · 0 papers · 40 scholar
Showing 49 of 49 items
KiCAD MCP is a Model Context Protocol (MCP) implementation that enables Large Language Models (LLMs) like Claude to directly interact with KiCAD for printed circuit board design.
AI Agents & Deep Learning: code, benchmarks and explainers. Building and operating agents on Microsoft Foundry and Azure; fine-tuning, inference optimization, precision and GPU benchmarks on NVIDIA and AMD; multimodal models.
SENEC.Home V2.x/V3/V4|P4 Integration for Home Assistant - successfully tested with latest V3: MCU v0842 / NPU v2608 (Sep 2026). Please consider sponsoring this project - TIA
SPICE notes on analog compute-in-memory circuits: simplified re-implementations of published macros on open PDK models, checking one claim per paper.
Rust/CUDA acceleration primitives for spiking neural networks and neuromorphic workloads, optimized first for NVIDIA Blackwell GPUs.
Neuromodulatory reward-shaping primitive for SNN plasticity systems: Environment → Critic → ModulatorVector (DA/5-HT/ACh/NE) for reward-modulated plasticity. Zero-dep, not an actor-critic or RL framework.
Design and FPGA implementation of a lightweight Spiking Neural Network (SNN) accelerator for energy-efficient inference, focusing on hardware-efficient neuron computation and neuromorphic processing.
BuildIt is a lightweight C++ library system for implementing domain specific languages (DSLs) and compilers through multi-stage execution. It uses partial evaluation to generate optimized code without requiring parsers, explicit intermediate representations (IRs), or custom code generators. History variables and forward execution support analyses of past computation, but optimizations such as data movement, operator fusion, and binding-time analysis also require information about future execution. We introduce prophecy variables into BuildIt to predict future program behavior and guide code generation. Repeated first-stage execution refines predictions until they satisfy the program's checks, extending BuildIt's existing reexecution machinery with little additional complexity. This approach obtains information typically computed by backwards dataflow analyses over an explicit IR while preserving BuildIt's lightweight implementation model. The cost of first-stage reexecution can be amortized over repeated executions of the generated code. We formalize the mechanism and mechanize its guarantees in Lean 4. We demonstrate prophecy variables in three applications: selective CPU-GPU data movement, neural-network operator fusion, and automatic binding-time analysis. These applications achieve performance improvements of up to 5.5x while simplifying the programming interface.
Longitudinal personal albums are weak-schema multimodal databases: noisy perceptual records whose key facts require joins across faces, text, timestamps, locations, and repeated events. Existing visual, video, document, and lifelog benchmarks test sub-problems, but not album-scale profile reconstruction with social identity binding and evidence citation. Benchmarking this task is difficult because the ground truth needed for evaluation--owner profiles, social graphs, face-name maps, and evidence provenance--is private state that real albums cannot safely release. We introduce PAL-Bench, a controlled benchmark for evidence-grounded reconstruction under a public-record contract. Its Evidence Compiler builds latent private worlds, programs target-level evidence paths, renders album pixels, re-measures them through perception pipelines, and exports audited public/private views. Agents receive only perception-derived public records; targets, identifier maps, and evidence paths remain hidden. PAL-Bench contains 50 synthetic users, 36,659 public photo records, and 2,799 targets over owner facts, identities, and relations. A privacy-preserving audit with 10 participants confirms that PAL-Bench evidence structures match real private albums, though equivalent releases remain privacy-prohibitive. Across seven systems and two compute-matched diagnostics, a seven-metric protocol reveals a gap between plausible profile summarization and faithful social reconstruction: systems recover some owner facts but struggle with recurring identities and evidence citation. PAL-TRACE, a reference framework that freezes identity bindings before owner-fact mining, performs best but leaves hard identity resolution far from solved. PAL-Bench provides a testbed for perceptual entity resolution, multimodal data integration, temporal evidence aggregation, and provenance-aware structured prediction.
GPUs accelerate a growing fraction of modern computing workloads. While prior work has studied the exploitation of memory errors in GPU applications written in C/C++ dialects, existing attacks rely on artificially introduced memory errors that grant attackers capabilities that may not arise when exploiting real-world GPU software. Consequently, GPU vendors and developers continue to largely treat such errors as reliability concerns rather than security issues, leaving vulnerabilities unpatched and potentially exploitable. This paper challenges that assumption. We show, for the first time, that remote, unprivileged adversaries can trigger device-side memory errors in realistic ML applications under specific conditions. We develop controlled proof-of-concept exploits that escalate vulnerabilities in the PyTorch ML framework and the CuDF dataframe library into write-what-where primitives, and weaponize such errors to enable denial-of-service, targeted manipulation and degradation of ML models, sensitive data leakage, and device-side code injection. To facilitate exploit construction, we design and implement a dynamic taint-tracking framework for NVIDIA GPUs that tracks attacker-controlled data flows to identify exploitable memory objects. Our findings demonstrate that adversaries can conduct memory-corruption attacks on vulnerable ML inference servers, highlighting the need to reassess prevailing assumptions about GPU memory safety and adopt essential mitigations, such as Write- XOR-eXecute, that are currently absent from modern GPUs.
Electric machines for aviation applications are currently being intensively developed. Increasing performance to 30 kW/kg at the 10–20 MW scale requires a step change in air-gap magnetic field and hence the wire current density. Fully superconducting machines greatly increase the current density in the motor windings and show promise for achieving the power-to-weight ratios required, and the use of liquid hydrogen as the aircraft propulsion energy source presents the opportunity to cool the superconductors with minimal additional refrigeration plant. In this article we simultaneously evaluate the superconductor AC losses and cryogenic cooling system requirements for two different armature conductors: a magnesium diboride (MgB2) winding at a nominal 20 K operating temperature, and a rare-earth barium copper oxide (REBCO) cable winding operating at a nominal 35 K. A 3-MW synchronous air-core motor model is used as the basis for comparison, in order to define relevant AC coil currents and alternating magnetic fields. For these initial comparisons, AC losses generated in the cables are estimated using electromagnetic finite element analysis (FEA), and heat transfer from the cable to the coolant is evaluated using a combination of thermal FEA and pipe-flow correlations. This in turn allows us to estimate coolant flows, temperatures, and pressures required in the aircraft cryogenic cooling system. We found that the performance of MgB2 exhibits a strong temperature dependence above 20 K which affects the physical dimensions of the cable as well as the losses generated, and this in turn impacts the heat transfer characteristics. The dependence on temperature is lower with REBCO cable, but losses are higher due to tape geometry. The two conductors lead to very different cooling system options and we can compare the advantages of each from an overall powertrain complexity and performance perspective.
CyHulu connects acoustic hulusi performance with real-time fingering visualization and an embodied digital musical companion. The system combines three interchangeable stopper configurations, finger-hole and breath sensing, and a browser-based interface featuring Alan, a three-dimensional companion. Its design builds on the instrument’s familiar body and playing technique while introducing visual guidance and accompaniment. Live sensor input supports fingering feedback, while score-driven animation enables Alan to demonstrate musical passages and perform a companion part. Our Real-Time Live! demonstration moves from instrument reconfiguration and live feedback to guided practice and duet performance, showing how physical instrument design and responsive graphics can support a shared musical experience.
PAIC, the patent-pending Promptable AI-Camera, is a software pipeline for real-time, prompt-driven transformation of live video. It is not a physical camera: it accepts feeds from cameras, webcams, smartphones, or rendered sources and routes the processed output into live production tools such as OBS. The creator uses prompts to define what should change, what should remain, and whether the original scene geometry should stay intact, be partly altered, or give way to another scene. The current implementation integrates FLUX klein as an edit model within a realtime frame-based video diffusion pipeline designed for low-latency, 24fps output. Its contribution for SIGGRAPH Asia Real-Time Live is the integration of four production features: realtime preview, temporal consistency, promptable control, and plug-and-play AI acceleration. The SIGGRAPH Asia Real-Time Live demonstration presents PAIC as a live production instrument in which backgrounds, regional edits, painterly or graphic visual languages, and rapid variations are generated while the subject remains on camera, turning image creation into something that can be performed during the live event.
Sharded blockchains enable scalable distributed ledgers by partitioning transaction workloads into multiple parallel processing shards. However, existing sharding protocols for heterogeneous blockchain environments mainly focus on mitigating performance degradation caused by uneven node bandwidth, while overlooking dynamic network variations, such as bandwidth fluctuations, unstable packet delivery, and transaction queue accumulation. These factors result in inefficient shard utilization and significant performance disparities among shards. To address these issues, we propose PabShard, an intelligent sharding framework for heterogeneous blockchain systems. PabShard leverages reinforcement learning to adaptively regulate shard operations according to network states and workload conditions. Specifically, we address two key challenges. First, to achieve effective shard organization under heterogeneous network conditions, we design a PPO-driven shard control mechanism that dynamically adjusts shard configurations based on observed system states. Meanwhile, an asymmetric bandwidth-adaptive asynchronous Merkle tree (BA-AMT) is introduced to improve transaction verification efficiency by exploiting available bandwidth resources. Second, to alleviate transaction congestion caused by heterogeneous shard processing capacities, we develop an adaptive block generation rate control mechanism that dynamically regulates the processing capability of individual shards. We implement PabShard on a Huawei Cloud cluster with heterogeneous bandwidth configurations ranging from 20 to 60 Mbit/s, where each physical host supports numerous virtualized blockchain nodes. Extensive experiments against representative sharding protocols, including Monoxide, GearBox, and Manifoldchain, demonstrate the effectiveness of PabShard. Under a transaction arrival rate of 3000 TPS with eight shards, PabShard achieves a 12.3% throughput improvement and reduces confirmation latency by 14.8% compared with the best-performing baseline.
BFT-like consensus protocols have two drawbacks: i) their three-phase synchronization latency is high, and ii) the communication overhead among consensus nodes is high. Consequently, we are motivated to design a lightweight consensus that enables low-communication asynchronous operations. To this end, we present a new consensus protocol, <monospace>Feather</monospace>, which is <italic>lightweight</italic> and enables <italic>asynchronous</italic> blockchains. <monospace>Feather</monospace> employs a two-phase <italic>propose-and-endorse</italic> consensus that uses a unique mechanism in which only a subset of nodes (i.e., proposers) participate in block generation asynchronously. Compared with BFT-like protocols, <monospace>Feather</monospace> can dramatically reduce the communication overhead and the state-synchronization latency. This is because <monospace>Feather</monospace> can achieve fast local approximate consensus across multiple asynchronous consensus zones, effectively eliminating forks. To demonstrate the practicality of <monospace>Feather</monospace> protocol, we implement a prototype using Golang and deploy it in two types of environments: i) a local deployment over 13 local high-performance physical machines, and ii) a global deployment of 100 cloud virtual machines across 10 global cities. The experimental results show that <monospace>Feather</monospace> achieves <inline-formula><tex-math notation="LaTeX">$\mathbf{2.5\times}$</tex-math><alternatives><mml:math><mml:mrow><mml:mn mathvariant="bold">2.5</mml:mn><mml:mo>×</mml:mo></mml:mrow></mml:math><inline-graphic xlink:href="huang-ieq1-3733340.gif"/></alternatives></inline-formula> and <inline-formula><tex-math notation="LaTeX">$\mathbf{7.3\times}$</tex-math><alternatives><mml:math><mml:mrow><mml:mn mathvariant="bold">7.3</mml:mn><mml:mo>×</mml:mo></mml:mrow></mml:math><inline-graphic xlink:href="huang-ieq2-3733340.gif"/></alternatives></inline-formula> higher transaction throughput than improved PBFT and Algorand.
Public smartphone charging is ubiquitous, yet poses a significant attack surface. Existing side-channel attacks have revealed various vulnerabilities but typically require line-of-sight access, installed malware, or tampered hardware, limiting their practicality. In this paper, we identify and exploit a novel physical vulnerability arising from common-mode leakage currents in ungrounded device chargers. We demonstrate that finger interactions on a smartphone touchscreen modulate this current by dynamically altering the capacitive coupling between the device and the user's body. These modulations encode fine-grained keystroke dynamics and can be captured without any modification to the victim device or charger. Capitalizing on this insight, we present a stealthy, non-intrusive keystroke inference attack that operates in a non-line-of-sight setting. By placing a concealed electrode (e.g., embedded in a seat cushion on which the victim sits), we capacitively intercept leakage signals through the user's body while the victim enters digits on a smartphone resting on a nearby surface. This setup decouples sensing from the target device and enables attacks even when the victim uses a trusted charger connected to an unmodified outlet. To address signal ambiguity, we design a physics-aware inference engine that fuses deep learning with explicit physical constraints. Specifically, we incorporate spatial-amplitude mapping and kinematic motion constraints to disambiguate spatially adjacent keys and correct sequence errors. Evaluated across diverse scenarios with 22 participants, our system achieves a Top-3 keystroke inference accuracy of 98.17%. Our system is device-agnostic and requires no victim-specific calibration, leveraging generic kinematic models to successfully recover 89.67% of 6-digit PINs within just 3 attempts. It also demonstrates the capability to reconstruct high-entropy inputs, underscoring the severity of this emerging physical threat.
A new approach to discretise and solve Richards equation based on the finite element method (FEM) is presented. The novelty of the approach, called the Neural Physics method, is that it expresses the discretisation as a convolution neural network (CNN) whose weights are predetermined by the integrals of the basis functions in the FEM. The Neural Physics method is physically consistent as it represents a discretisation of the governing equation and can run on CPUs, GPUs and the latest AI accelerators, resulting in an increase in computational speed. The Neural Physics method is demonstrated for two different discretisations of Richards equation, including a multigrid method that is realised through a U-Net architecture. The accuracy and speed of these methods is tested on a problem with a known analytical solution and compared to a standard FEM, and shows that the Neural Physics method is almost an order of magnitude faster than a comparable FEM. Finally, the Neural Physics method is applied to a real-world test case with integrated experimental data and simulates the infiltration of water through a rain garden. This new approach therefore provides a route to efficiently and accurately solve Richards equation for a wide array of hydrological applications.
Modern deep neural networks (DNNs) are prohibitively costly to build, making model intellectual property (IP) protection critical in computing systems. Trusted execution environments (TEEs) provide hardware-enforced data protection and are widely supported on CPUs, but comparable support remains limited on DNN accelerators, including GPUs. Existing TEE-based model-protection schemes improve inference performance by outsourcing computation from CPU TEEs to untrusted accelerators. However, their runtime and offline security-protocol overheads leave a substantial performance gap relative to native accelerator execution. This paper presents DeepIvy, an end-to-end system framework that protects model confidentiality while enabling efficient CPU-accelerator inference using CPU-side TEE support. DeepIvy employs privacy-aware model construction to train a unidirectional private adapter (UPA) attached to a frozen public model. The lightweight UPA is shielded inside the CPU enclave and receives features only from the public model executing on an untrusted accelerator. This design uniquely addresses key performance and security challenges of prior TEE-outsourcing techniques: (i) its unidirectional feature flow removes back-and-forth layer dependencies across devices, unleashing high CPU-GPU parallelism; and (ii) because transferred features originate only from the public model, DeepIvy eliminates input/output masking for TEE-resident layers and avoids costly offline mask computation and storage. DeepIvy further combines TEE-aware adapter search (TAS), which minimizes the UPA, with a system-level zero-copy technique that reduces tensor-copy overhead across enclave boundaries. We implement DeepIvy on an SGX2-based server and evaluate it on representative vision and NLP models. DeepIvy achieves TEE-GPU inference speedups of up to 22.2× and 9.8× over two state-of-the-art schemes while retaining up to 87.5% and 90.7% of native GPU-Only throughput for vision and NLP models, respectively. It provides strong hardware-backed model-weight protection and exhibits black-box-level robustness against query-based model-extraction attacks. Our work highlights the promise of accelerator-speed secure inference through model-architecture and system co-design.
Neural-symbolic AI (NSAI) combines learned perception with programs, constraints, and compositional reasoning, but remains difficult to test because small semantics-preserving changes can alter grounded symbols and trigger discontinuous reasoning failures. Existing neural fuzzers mainly rely on neural coverage or representation-space feedback and do not provide a common testing interface for heterogeneous NSAI systems. We introduce NSAI-Fuzz, a unified semantics-aware fuzzing framework that separates task-specific valid mutation from architecture-neutral guided search and represents each candidate using five availability-aware, rank-normalized risk signals: boundary pressure, grounding uncertainty, representation drift, constraint violation, and archive novelty. We further propose Reliability-Calibrated Weighting (RCW), which learns dataset-model-level signal allocations from calibrated reliability, and Contextual Signal Gating (CSG), which adapts signal weights to the current candidate pool. We evaluate four NSAI systems across four reasoning tasks, nine compatible dataset-model cells, 20 policies, and 1,800 on-policy runs. Fixed five-signal guidance outperforms Random and Uniform in every cell, while RCW and CSG further improve over Fixed in all nine cells. Under matched mutation and query budgets, RCW and CSG achieve 16.17% and 16.06% aggregate failure discovery, compared with 6.06% for the strongest neural-fuzzing baseline. Wall-clock evaluation further shows that TensorFuzz requires 1.11x the runtime of RCW while achieving only 0.45x its failure throughput. These results show that effective and efficient fuzzing of heterogeneous NSAI systems benefits from a common neural-symbolic risk interface with adaptive signal allocation.
Neural architecture search (NAS) can automate medical image segmentation network design, but costly candidate evaluation limits the supervision available for evolutionary selection. Since replacement decisions depend on relative candidate quality, learning architecture preferences directly can provide guidance under constrained training budgets. In this paper, we propose Ranking-Aware Predictor-Assisted Differential Evolution (R-PADE) for U-Net-style architecture search. Specifically, a predictor trained with pairwise ranking supervision and contrastive regularization directly determines trial–target replacement in discrete differential evolution. Aggregated pairwise preferences prioritize candidates for proxy training, whose observations subsequently update the predictor. This connects evolutionary selection with selective candidate evaluation, reducing reliance on direct training for replacement decisions. Structured encoding and feasibility repair maintain valid architectures during the search. Experiments on CHAOS, BUSI, and ISIC demonstrate competitive segmentation performance with compact architectures. Compared with EU-Net, R-PADE reduces search GPU time by 35.6% while achieving higher Dice scores on CHAOS and BUSI. The ablation experiments compare the implemented predictor formulations and evaluate the contribution of contrastive regularization using matched search settings and proxy-evaluation budgets.
Like many multiplayer online battle arena games, Valve Software's Dota 2 begins with a crucial stage of alternating strategic bans and selections ("picks") of heroes to play in the game. We present a framework for automating these ban/pick choices which first trains an oracle to predict game outcomes given a full set of picks from both team, and then uses one of several methods to make interleaved ban/pick actions based on the oracle. These methods apply Monte Carlo Tree Search, coevolutionary optimization, and transformer neural network architectures. We test these methods against one another, and compare them, and ensembles of them, to the existing literature.
Methyltransferases (MTs) are a functionally diverse class of enzymes that catalyse methyl group transfer from S-adenosyl-L-methionine (SAM) to DNA, RNA, and proteins, thereby regulating epigenetic programming, gene expression, and numerous disease processes. Dysregulation of these enzymes is implicated in a wide spectrum of human cancers and neurodegenerative disorders, making their accurate computational identification a priority in both basic research and drug discovery. We present GAC-BiT-CNN, a novel hybrid deep learning framework integrating Generative Adversarial Capsule Networks (GAC), Bidirectional Transformers (BiT), and Convolutional Neural Networks (CNN). The model leverages high-dimensional protein embeddings from four protein language models, including ProtGen, ESM-2, ProtT5-BFD, and ProtT5-XL, used individually, together with a fused feature set formed by concatenating the ProtT5-BFD and ProtT5-XL embeddings. All architectural components were optimised via systematic grid search hyperparameter tuning, and SHapley Additive exPlanations (SHAP)-based analysis was applied to the fused embeddings to identify the most discriminative feature dimensions. GAC-BiT-CNN with the fused feature set achieved 98.34% accuracy on the training set and 95.50% accuracy on the held-out test set, with an area under the receiver operating characteristic curve (AUC) of 98.70% and a Matthews correlation coefficient (MCC) of 0.901, surpassing all baseline models by a substantial margin. GAC-BiT-CNN provides a scalable, alignment-free framework with representation-level SHAP attribution for high-throughput methyltransferase identification, with direct applications in epigenomics, drug discovery, and cancer biology.
Dense optical flow allows small drones to avoid obstacles. However, few dense optical flow networks are designed for microcontrollers (MCUs) with neural accelerators, whose on-chip SRAM limits the input resolution. We present PACE-NAS, a macro-to-micro neural architecture search framework that decouples memory from latency: a macro design fixes the backbone and its peak SRAM, and a micro design searches operators and depths to trade accuracy against compiler-estimated latency. We apply it to EdgeFlowNet on a Grove Vision AI V2 board (Ethos-U55), where profiling locates the peak SRAM at the full-resolution output head. Replacing transposed convolutions with bilinear upsampling and convolutions lowers the peak SRAM by 18%, and two lightweight context modules lower the endpoint error (EPE) without raising the peak. All searched subnets share this peak, so the search needs no memory constraint. We name the Pareto-optimal models MCUFlowNet; the fastest variant, MCUFlowNet-S, runs 1.76x faster than a retrained EdgeFlowNet at a 1.6% higher EPE and, in simulated forest flights at 2 m/s, reaches the goal without collision in 51.5% of trials, against 41.8%. PACE-NAS also applies to semantic segmentation, where a searched PIDNet-S variant runs 1.17x faster at similar accuracy. Code is available at https://gitlab.kuleuven.be/emedia-research-lab/esat-stadius/mcuflownet/mcuflownet.
Natural and man-made disasters require rapid visual assessment to support emergency response and resource allocation. Unmanned Aerial Vehicles (UAVs) can capture aerial imagery from disaster-affected regions, but practical deployment requires compact deep learning models that maintain strong classification performance under limited memory and model-storage constraints. This paper presents a parameter-constrained Neural Architecture Search (NAS) approach for lightweight UAV disaster image classification. The search space is built from mobile inverted residual blocks and optimized using an evolutionary strategy with proxy training. The selected NAS architecture contains only 51,661 trainable parameters and is evaluated on a five-class disaster image dataset containing earthquake, fire, flood, hurricane, and normal scenes. We also evaluate an EfficientNet-B0 weight-transfer variant and compare it with the same NAS architecture trained from scratch. Across three random seeds, the scratch-trained NAS model achieves 96.34% ± 0.51% accuracy and 96.30% ± 0.51% macro-F1. The weight-transfer variant achieves 95.27% ± 0.68% accuracy, showing that direct layer-order transfer is not consistently beneficial for this compact architecture. Larger pretrained baselines achieve higher absolute accuracy but require substantially more parameters. The results show that the proposed NAS architecture provides a compact parameter-efficient alternative for UAV-oriented disaster classification.
Quantum cloud providers can read the circuits they receive and their outputs, yet methods for hiding a circuit are usually evaluated on the circuit as written, before compilation optimises it. We compare three ObfusQate schemes (inverse, composite and delayed gates), ECQCO and Q-MaskCut under one protocol that compiles every protected circuit at optimisation level 3 before measuring it; the goal is to keep the provider from learning the user's algorithm and answer. Q-MaskCut is one example combination of existing techniques: a one-time-pad key, decoy wires, circuit cutting with a random rotation pair across each cut, and Pauli twirling with dynamical-decoupling masking after compilation. We simulate 11 circuits of 3 to 19 qubits without noise and on two noise models, measure the compiled structure and execution time of 23 circuits of 3 to 127 qubits, and test whether the structural disguise survives a second compilation by the provider. Composite and delayed gates lose most of their structural disguise in compilation and leave the output readable; inverse gates keep a high normalised graph edit distance (normGED) only by making execution 7 to 112 times longer; ECQCO's output could not be decoded with the implementation provided by its authors. Q-MaskCut hides the output (mean total variation distance without the key 0.94), and without noise, with exact fragment probabilities, it decodes exactly. Its compiled normGED of 0.88 comes mostly from the masking, which a second compilation removes, leaving 0.42 from the key, decoys and padlocks. With noise, its fidelity is comparable to the original's; it needs 7 to 91 fragment variants per circuit in the simulations (7 to 61 in a run on ibm_marrakesh) and a classical reconstruction. Protection should therefore be evaluated on the compiled circuit and again after the provider recompiles it.
Security decisions about a web page are best made where the page is: in the browser, before anything leaves the machine. We present the Browser Immune System (BIS), a browser extension that screens navigation with a cybersecurity transformer running entirely on the user's GPU, and alias-web, the model that powers it: a 4B-parameter model that starts from the dual-regime cybersecurity transformer (CT0), extends its public release alias3-micro with a context-category sensor, and is quantised to 4 bits and compiled to a 2.44 GB graph for WebGPU. alias-web is the first in-browser realisation of that model's typed decision mode, System 1, with its official pieces: the marker-token mapping, the complete pointer head and the final normalisation that the exported graph omits. It answers a typed security question in about 150 ms on a laptop GPU, and the same weights generate explanations at 10.5 tokens/s on demand. BIS treats the model as an immune system: an innate response, one System 1 question per threat class, runs on every page; an adaptive response, System 2 text, explains a verdict only when the user asks. On BISBENCH-8, eight synthetic sites of which six execute real but harmless malicious behaviour and two are adversarial controls, BIS returns the expected page verdict on 7/8 sites with both controls green and a class-level micro-F1 of 0.71, scans a page in 0.9 to 6.4 s, and gives identical verdicts across samples and across Chrome and Firefox. Single-variable diagnostics trace both benchmark errors to what the model was shown and, with a first pass over eleven real public pages, show that two of the seven questions answer yes on ordinary scripts and login forms: specificity on the real web is the open problem. We report every error, the engineering obstacles between a research checkpoint and a working extension, and what the evidence does not yet support.
Cultural heritage sites are increasingly monitored by sensor networks, but querying their data requires database expertise. Large Language Models (LLMs) offer a natural-language interface, yet over raw sensor records they lose track of which sensor, site, and time each value belongs to, and silently miscompute. We present HeriKGQA, a knowledge-graph question-answering method that confines the LLM to interpreting questions and phrasing answers, leaving computation to SPARQL and deterministic code. Questions are parsed into intents constrained by a heritage ontology aligning W3C SOSA/SSN with CIDOC-CRM, compiled into SPARQL, and answered from query results and temporal, spatial, and anomaly facts derived by a grounding layer, with error correction guarding each LLM step. On 500 questions over real data from five heritage sites, HeriKGQA reaches 99.0% accuracy, 18.2 points above the strongest of three baselines using the same LLM and receiving the same anomaly rules and answer conventions, with comparable answer quality while running almost four times faster. Without grounding, accuracy drops by 12.2 points and on anomaly questions from 99% to 66%. We report lessons from development and live deployment.
Liquid-to-gas phase-change actuators provide compact, compliant actuation, but their nonlinear pressure dynamics, thermal delay, and asymmetric heating-cooling response make precise control difficult. Building on an established Physical Hysteretic Reservoir Computing (PhysRes) architecture, this letter integrates a two-model PhysRes predictor with sampling-based Model Predictive Control (MPC) and deploys it on an edge platform. Separately trained scalar pressure and joint two-output temperature models use shifted-sigmoid reservoirs, with only their linear readouts fitted by closed-form ridge regression. With 1,050 training samples, PhysRes predicts chamber pressure, Peltier temperature, and liquid temperature with <inline-formula><tex-math notation="LaTeX">$R^{2} \geq 0.9977$</tex-math></inline-formula> for each state. Among all evaluated predictors, PhysRes achieves the lowest measured desktop one-step inference latency of <inline-formula><tex-math notation="LaTeX">$0.013 \,\mathrm{m}\mathrm{s}$</tex-math></inline-formula>, 12.0× faster than a fully trained long short-term memory (FT-LSTM) network. Across experiments at 120, 130, and <inline-formula><tex-math notation="LaTeX">$140 \,\mathrm{k}\mathrm{Pa}$</tex-math></inline-formula>, PhysRes-MPC achieves a mean absolute error (MAE) of <inline-formula><tex-math notation="LaTeX">$1.95 \,\mathrm{k}\mathrm{Pa}$</tex-math></inline-formula>, <inline-formula><tex-math notation="LaTeX">$32.2 \,\%$</tex-math></inline-formula> lower than proportional-integral-derivative (PID) control, and the lowest maximum overshoot at each target. On the ESP32, frozen PhysRes and FT-LSTM pressure predictors require 157.0 and <inline-formula><tex-math notation="LaTeX">$2090.2 \,\mathrm{\upmu }\mathrm{s}$</tex-math></inline-formula> per inference, respectively; their corresponding 200-candidate, 10-step rollouts require 316.5 and <inline-formula><tex-math notation="LaTeX">$4205.5 \,\mathrm{m}\mathrm{s}$</tex-math></inline-formula>.
Superconducting digital (SCD) technology has long carried promise for advancing large-scale computing leveraging the unique physical characteristics of superconducting materials and devices, such as low-voltage, high-frequency operation and zero-resistance interconnects. At the same time, SCD technology has also historically been beset with many scaling challenges. However, recent critical advances including improved power delivery, high integration density and availability of superconducting random-access memory prompt a fresh evaluation of SCD-based computing. In this paper, we study whether SCD systems can outperform conventional CMOS and cryo-CMOS for large-scale workloads even after accounting for full system overheads. Using a novel multilevel analytical modeling framework, our study explores performance implications of various SCD configurations across a range of architecture and technology assumptions in applications such as large language models (LLMs). The results show that, despite cooling overheads, a key benefit of recently proposed SCD technology is 9<inline-formula><tex-math notation="LaTeX">$\boldsymbol{\times}$</tex-math><alternatives><mml:math><mml:mo mathvariant="bold">×</mml:mo></mml:math><inline-graphic xlink:href="peraltavelazquez-ieq1-3740058.gif"/></alternatives></inline-formula> - 286<inline-formula><tex-math notation="LaTeX">$\boldsymbol{\times}$</tex-math><alternatives><mml:math><mml:mo mathvariant="bold">×</mml:mo></mml:math><inline-graphic xlink:href="peraltavelazquez-ieq2-3740058.gif"/></alternatives></inline-formula> better energy efficiency. By contrast, performance gains ranging between 4<inline-formula><tex-math notation="LaTeX">$\boldsymbol{\times}$</tex-math><alternatives><mml:math><mml:mo mathvariant="bold">×</mml:mo></mml:math><inline-graphic xlink:href="peraltavelazquez-ieq3-3740058.gif"/></alternatives></inline-formula> - 16<inline-formula><tex-math notation="LaTeX">$\boldsymbol{\times}$</tex-math><alternatives><mml:math><mml:mo mathvariant="bold">×</mml:mo></mml:math><inline-graphic xlink:href="peraltavelazquez-ieq4-3740058.gif"/></alternatives></inline-formula> are more limited and strongly depend on the availability of high-density, high-bandwidth on-chip memory that traditionally was not attainable. In general, the results show that recent technological advances are key to a promising scalable path forward, where our ablation studies highlight critical aspects of the maturity of SCD technology.
Integration of Silicon-Photonic (SiPh) interconnects with multi-chiplet systems offers unprecedented bandwidth and energy efficiency for next-generation computing. However, the resulting high power density and complex vertical thermal coupling create severe thermal challenges that threaten both computational integrity and photonic communication reliability. Accurate and efficient full-chip thermal monitoring is a critical prerequisite for dynamic thermal management, yet it remains challenging. The limited number of physical thermal sensors that are often sparsely placed on the compute chiplets fails to capture the intricate 3D thermal landscape, especially within the photonic interposer network (PIN) itself. To this end, we propose TherMoPIN, a novel framework achieving accurate yet cost-effective full-chip thermal monitoring. In addition to the physical thermal sensors that are readily placed on chiplets, it repurposes the existing PIN into a uniformly placed virtual thermal sensor array with negligible cost increases. TherMoPIN comprises two key enablers. First, we propose ArbLink, a precise and cost-efficient thermal sensing mechanism, which transforms the entire PIN into a dense and uniformly-distributed ‘virtual’ sensor array by reusing available photonic routers for thermal sensing, with minimal hardware overhead. Second, we develop HorVer, a lightweight thermal reconstruction technique that fuses sparse, direct readings from horizontal on-chiplet thermal sensors (i.e., physical ones) with the dense, indirect thermal data from the vertical photonic interposer network (i.e., virtual ones), capturing and modeling the critical thermal coupling between the compute chiplets and the photonic interposer. This unique horizontal-vertical data fusion allows HorVer to accurately reconstruct the full-chip thermal map of SiPh-chiplet systems. Experimental results demonstrate that TherMoPIN achieves high thermal monitoring accuracy, significantly outperforming state-of-the-art (SOTA) methods that rely solely on physical sensors. Furthermore, it accomplishes this with negligible additional hardware cost, as it primarily reuses the already-present photonic infrastructure for sensing.
Generative large language models (LLMs) face significant challenges in deployment due to their substantial computational and memory demands, particularly under batched inference scenarios. While activation sparsity offers a promising direction to reduce resource consumption, its benefits diminish with increasing batch sizes. Moreover, existing sparse inference systems fail to fully utilize modern GPU Tensor Cores. This paper proposes <italic>SpBatch</italic>, an efficient execution framework for batched activation-sparse FFN computation in ReLU-adapted LLMs through a two-level design. At the system level, it introduces <italic>Sparsity-Aware-Division</italic> to dynamically partition large batches into smaller mini-batches, maximizing sparsity and hardware utilization simultaneously. At the kernel level, it employs a “<italic>Load by CUDA Cores, Compute by Tensor Cores</italic>” strategy, alongside a <italic>Bitmap-based Neuron Index</italic> mechanism for load balancing and an optimized execution pipeline. SpBatch achieves up to 8.89<inline-formula><tex-math notation="LaTeX">${\boldsymbol{\times}}$</tex-math><alternatives><mml:math><mml:mrow><mml:mo mathvariant="bold">×</mml:mo></mml:mrow></mml:math><inline-graphic xlink:href="zhao-ieq1-3732029.gif"/></alternatives></inline-formula> speedup over PowerInfer (a representative activation-sparse baseline) and up to 2.6<inline-formula><tex-math notation="LaTeX">${\boldsymbol{\times}}$</tex-math><alternatives><mml:math><mml:mrow><mml:mo mathvariant="bold">×</mml:mo></mml:mrow></mml:math><inline-graphic xlink:href="zhao-ieq2-3732029.gif"/></alternatives></inline-formula> over llama.cpp (a dense inference framework) in end-to-end inference under the evaluated settings. These results demonstrate that activation sparsity and batch processing can be jointly leveraged through a selective sparse FFN execution path, especially in small-to-moderate batch regimes and high-sparsity edge-oriented scenarios where dense execution does not yet fully dominate.
Confidential computing aims to protect tenants in cloud computing settings against malicious privileged software controlled by the cloud provider. AMD SEV-SNP is one approach to enable confidential computing, by providing guest virtual machines with confidentiality and integrity of data in use. AMD uses an on-SoC hardware component called Platform Security Processor (PSP) to implement parts of SEV-SNP. We present a new attack Mior, to show that a malicious software adversary with hypervisor privileges (e.g., cloud provider) can misconfigure the platform to change the memory view of the PSP. Specifically, Mior coerces the PSP into reading arbitrary data in our control, and redirects PSP writes to our attacker-controlled destination. In other words, Mior creates an attacker-controlled memory for the PSP. We refer to this primitive as Device-as-Memory. To showcase the impact of Mior, we use the Device-as-Memory primitive to corrupt PSP writes during the SEV-SNP lifecycle (e.g., initialization, attestation), resulting in violation of all security guarantees afforded by SEV-SNP.
Once specialized for graphics acceleration, Graphics Processing Units (GPUs) now serve as foundational components of parallel computing, driving advancements from immersive gaming to generative Artificial Intelligence (AI). However, while their computational capabilities have expanded dramatically, GPU security has lagged behind their expanding functionality. Recent studies reveal that GPUs remain susceptible to memory corruption attacks, a gap amplified by their reliance on memory-unsafe programming frameworks, such as CUDA. Although contemporary GPUs incorporate basic defenses like ASLR and stack canaries, the absence of hardware-enforced Data Execution Prevention (DEP) leaves them critically exposed to code-injection attacks, undermining the implementation of a robust W⊕X security policy. This work analyzes the potential of code-injection attacks on NVIDIA GPUs, executed independently of CPU vulnerabilities, in the presence of ASLR and stack canary protections. To address this gap, we design and implement a software-based DEP mechanism, a critical enabler of the W⊕X policy, to enforce strict memory permissions, eliminating executable-writable memory regions, and thereby neutralizing code-injection attacks. Our solution operates transparently alongside proprietary vendor toolchains and software components while maintaining compatibility with precompiled libraries. Empirical evaluations across diverse benchmarks on three different NVIDIA GPU architectures show that our approach incurs minimal overhead, with average runtime and memory overheads of less than 0.5%.
GPUs accelerate a growing fraction of modern computing workloads. Yet, modern GPU programming frameworks, including CUDA, ROCm, and Triton, expose unchecked pointer arithmetic, leaving GPU kernels vulnerable to the same spatial memory errors as C and C++ programs. Despite prior work studying the exploitation of memory errors in GPU applications, GPU vendors and developers have not been convinced and consequently continue to largely treat such errors as reliability concerns rather than security issues, leaving vulnerabilities unpatched and potentially exploitable. During the first two years of this PhD, we have challenged this assumption. In particular, we have demonstrated that remote, unprivileged adversaries can trigger device-side memory errors and, under specific conditions, escalate them into write-what-where primitives, enabling denial-of-service, targeted manipulation and degradation of ML models, sensitive data leakage, and device-side code injection. At the same time, we notice that existing GPU memory-safety tools have seen limited adoption as they impose high overhead, require hardware changes, or fail to integrate with production GPU software stacks. Therefore, we propose three lightweight memory-safety mitigations, based on insights from our offensive research, tailored to the strengths of GPUs and their deployments. Specifically, we discuss continuous fine-grained memory-layout re-randomization, data-space randomization, and Right-on-Time Safety (RoTS): a software-only memory-safety mitigation that validates GPU memory-access ranges at the host-side kernel launch whenever those ranges depend only on launch-time values, and emits device-side checks only for data-dependent accesses.
The rise of edge-centric 6G networks demands real-time intelligent processing of heterogeneous radio-frequency (RF) signals. Complex-valued blind source separation (CV-BSS) provides a physical-layer capability, enabling the blind extraction of source components from overlapping RF signals. Here, we demonstrate an integrated photonic vector processor (IPVP) that implements real-time CV-BSS via in-physics computation directly on raw RF signals. The IPVP operates across multiple RF bands up to 24 GHz and achieves blind, in-hardware adaptation to unknown sources and channels via in situ learning. Experiments demonstrate the efficacy of IPVP in multiple-input multiple-output (MIMO) communications and multi-band radar interference suppression. The IPVP reduces the EVMs of the two recovered signals from over 50% to 6.18% and 5.3%, respectively, while enhancing the SIR by up to 26 dB. This work demonstrates, to our knowledge, the first photonic processor capable of real-time CV-BSS at RF, highlighting the potential of integrated photonic hardware for ultralow-latency, edge-intelligent signal processing.
Quantum natural language processing (QNLP) turns the grammar of a sentence into a tensor network and then into a quantum circuit with trainable angles. We build such a model for molecules: each heavy atom is a word with one wire per bond, and each bond is a ``cup''joining two wires, so the circuit's wiring is the molecular graph. We work out exactly what these circuits keep. With the typed angle sharing of lambeq, the circuits separate every isomer pair we tested. Yet for isomers that simpler sharing confuses, the output states stay almost the same, before and after training. On 7 Tox21 pairs and 1,988 MOSES pairs, nearly all the difference lies in how often the measurements succeed, which a standard readout ignores. Simpler sharing, by element and bond type, shows why: the output then depends only on a graph whose vertices are the bonds. This explains all but one of the identical-output pairs we certify, including 1.19\% of 232{,}191 MOSES molecules, and for alkyl side chains it reduces to a tree shape that tells chains apart up to 15 atoms but not from 16. On small pairs, this circuit and the 1-WL graph test each separate pairs the other misses. We also show that SMILES writing order can change the output, that a run succeeds with probability at most $2^{-n}$ for $n$ bonds, and that a cup-free design avoids this cost. No circuit beats a simple atom-count model on Tox21, and we claim no quantum advantage.
Existing on-device AI architectures for resource-constrained environments face two critical limitations: they lack compactness, with parameter requirements scaling proportionally to task complexity, and they exhibit poor generalizability, performing effectively only on specific application domains (e.g., models designed for regression tasks cannot adapt to natural language processing (NLP) applications). In this paper, we propose CURA, an architecture inspired by analog audio signal processing circuit that provides a compact and lightweight solution for diverse machine learning tasks across multiple domains. Our architecture offers three key advantages over existing approaches: 1) Compactness: it requires significantly fewer parameters regardless of task complexity; 2) Generalizability: it adapts seamlessly across regression, classification, complex NLP and computer vision tasks; and 3) Complex pattern recognition: it can capture intricate data patterns while maintaining extremely low model complexity. We evaluated CURA across diverse datasets and domains. For compactness, it achieved equivalent accuracy using up to 2,500 times fewer parameters compared to baseline models. For generalizability, it demonstrated consistent performance across four NLP benchmarks and one computer vision dataset, nearly matching specialized existing models (achieving F1-scores up to 90%). Lastly, it delivers superior forecasting accuracy for complex patterns, achieving 1.6 times lower mean absolute error and 2.1 times lower mean squared error than competing models.
Overcurrent Protection (OCP) is a critical safety mechanism in Printed Circuit Board (PCB) design. Early-stage OCP verification remains largely manual because datasheet equations, pin settings, component ratings, schematic connectivity, and partfiles must be interpreted jointly. This paper presents a Large Language Model (LLM)-driven framework that combines datasheet-native Retrieval-Augmented Generation (RAG), Portable Document Format (PDF) text reconstruction, topology-guided parameter resolution, and deterministic rule checking. It extracts OCP specifications, computes tolerance-aware current-limit bounds, and checks protection rules before physical implementation. Experiments on real PCB designs demonstrate accurate automated verification while reducing manual analysis effort and late-stage iteration risk.