Research
Academic papers, new techniques, benchmarks, and theoretical findings in AI/LLM security.
Academic papers, new techniques, benchmarks, and theoretical findings in AI/LLM security.
2,033 items
The paper proposes K-TCDP, a finite-memory noise mechanism for differentially private LoRA supervised fine-tuning of large language models. It adds controlled negative temporal correlation to full-dimensional Gaussian noise in the LoRA adapter, so earlier perturbation can be partly offset later, while remaining compatible with standard RDP accounting. On four GLUE tasks and DART (BLEU-4, ROUGE-L), it consistently outperforms DP-SGD at the same privacy levels with small computational overhead.
Fed-C&E is a federated learning method for unsupervised skeleton-based action recognition that uses a closed-loop condensation-expansion paradigm. Clients condense their data, and the server expands it through a dual-level mechanism that synthesizes samples with a prototype-to-sequence similarity transformation matrix pool and supplements global information using second-order client statistics. The authors report that it outperforms aggregate-then-adapt FL methods across multiple datasets while preserving data privacy.
AmbShield is a physical layer security scheme that uses naturally distributed ambient backscatter devices (AmBDs) to protect wireless links against passive eavesdroppers without extra transmit power. The AmBDs act as friendly jammers that randomly backscatter to interfere with eavesdroppers and as passive relays that backscatter the desired signal to boost legitimate receivers. The authors derive closed-form secrecy outage probability and secrecy diversity order, and validate the gains with Monte Carlo simulations under imperfect synchronization and CSI estimation.
BlockAthena is a forensic framework for analyzing long-term on-chain crimes in account-based blockchains. It segments long-term transaction topology into subgraphs based on crime evolution periods, modeling direct and co-occurrence transactional behaviors. Experiments report an average 18% improvement in F1-score and up to an 80% reduction in memory overhead compared to the best-performing baseline.
This paper presents an optimal visual time-of-flight imaging scheme that learns iToF coding functions and depth reconstruction end to end, guided by Fisher information supervision. It adds RGB-derived vision information and a dual-branch reconstruction network with edge guidance to improve depth accuracy under low signal-to-noise conditions. The authors report effectiveness on synthetic and real-world datasets.
The paper introduces DFedCata, a decentralized federated learning algorithm that applies Catalyst Acceleration to address slow convergence and poor generalization caused by data heterogeneity. It combines a Moreau envelope function, which reduces parameter inconsistencies among clients, with a Nesterov extrapolation step that speeds up aggregation. The authors report improved convergence speed, computational cost and generalization on CIFAR10/100 and Tiny-ImageNet under various non-iid data distributions.
Multi-view clustering with joint training can leave some views under-optimized, because one view with more discriminative information dominates learning. The authors analyze this imbalance through gradient descent on each view-specific feature extractor and propose balanced multi-view clustering (BMvC), which adds a view-specific contrastive regularization (VCR) to modulate each view's optimization. BMvC is reported to outperform state-of-the-art methods on eight benchmark datasets and two spatially resolved transcriptomics datasets.
Researchers analyze the physical layer security and covert communication of an autonomous aerial vehicle (AAV)-assisted free-space optical (FSO) system that uses downlink rate splitting multiple access (RSMA) against an eavesdropper named Eve/Willie. The analysis models the generalized Málaga atmospheric turbulence, fog, boresight pointing errors and angle-of-arrival fluctuations, and derives series-based expressions for detection error probability, covert rate, generalized secrecy outage probability, average fractional equivocation and average information leakage rate. The authors optimize rate parameters to minimize secrecy outage and maximize confidential-message throughput.
Researchers present the first formal definition of honeypots arising from high-frequency interactions with decentralized applications (DApps). Because transaction records and pending transactions are publicly visible on blockchains, attackers can predict a user's parameters and build honeypots that trick the user into actions. Using a log-based transaction parsing method, the study finds 99 smart contracts affected by 636 incidents, with losses exceeding 25M USD, and a frequent mining method recovers the causes of victim transaction reverts with an 82.39% success rate.
Bi3D++ is a hybrid bi-domain active learning strategy for cross-domain 3D object detection in autonomous driving. It selects source data resembling the target domain and picks a small set of target point clouds for annotation, using rare-class similarity and intra- and inter-frame diversity. Across cross-beam and cross-location settings, it outperforms state-of-the-art UDA methods with only 1% labeled target data.
The paper proposes AFM–CIR, a framework for domain generalization in time series that combines semantic-similarity-guided Adaptive Fourier Mixing (AFM) with Causality-Inspired Regression (CIR), targeting remaining useful life (RUL) prediction under the strict DG setting where no target domain data is available during training. The authors report that AFM–CIR consistently achieves state-of-the-art performance across four widely used benchmark industrial data sets, outperforming ERM, general DG, and task-specific DG baselines.
This ACM Computing Surveys article, published in Volume 58, Issue 11 (pages 1-35, August 2026), surveys algorithm debt in machine and deep learning systems. The source text provided here contains only the bibliographic citation and no article body, so the defining, smell-cataloguing and future-work content it names cannot be described.
This paper proposes an explicit polar coding scheme for a two-user discrete memoryless multiple access wiretap channel with partial rate-limited feedback (MAC-WT-PLF). The feedback creates correlation between channel inputs, which existing polar schemes assuming independent inputs cannot support, and the scheme combines source polarization, block Markov coding, superposition coding, lossy source coding and rate-splitting. Asymptotic analysis establishes that it ensures reliability and strong secrecy and attains the entire achievable secrecy rate region from prior work.
The authors present a unified framework for interpreting vision models that centers on the pointwise feature vector (PFV) paired with its instance-specific Effective Receptive Field (iERF). It combines Sharing Ratio Decomposition (SRD) for local saliency maps, Concept-Anchored Feature Explanation (CAFE) for global semantic labeling of sparse autoencoder latents, and Interlayer Concept Attribution (ICAT) for tracing concept influence across layers. Across ResNet50, VGG16, and ViTs, the framework is reported to outperform baselines in fidelity and robustness.
The OWASP GenAI Security Project has launched FinBot, a hands-on Agentic Security capture-the-flag platform that is part of its Agentic Security Initiative. FinBot simulates a multi-agent vendor management platform with autonomous onboarding, fraud detection, invoice processing and communications, all powered by LLMs with real tool access. Its challenges cover prompt injection, tool misuse, policy bypass, data exfiltration, privilege escalation and remote code execution, mapped to the OWASP Top 10 for LLM Applications, the OWASP Top 10 for Agentic Applications, CWE and MITRE ATLAS.
The authors propose a graph-based defense model for whole slide images (WSIs) used in cancer diagnosis. It combines a graph neural network, a denoising module based on untrained neural network prior and graph signal processing, and a vision transformer for prostate cancer Gleason grading. Tested against various adversarial attacks, the model showed a significant improvement in diagnosis accuracy.
Fix: Based on the causes that lead to transaction reverts, the authors propose potential strategies to mitigate these security risks and validate them in a simulated environment. The specific strategies are not described in the source text.
IEEE Xplore (Security & AI Journals)This paper addresses fairness in counterfactual explanations (CFs), which show how input changes would alter an ML model's decision and so offer actionable recourse. The authors define individual, group and hybrid fairness, and propose a model-agnostic, reinforcement learning based method that generates CFs meeting fairness constraints at both individual and group levels. Evaluated on four benchmark datasets, the approach preserves proximity and plausibility of the generated CFs.
LVMark is a watermarking method for latent video diffusion models that aims to protect model ownership. It uses a watermark decoder that learns consistency between adjacent frames, combining low-frequency components of the three-dimensional wavelet domain with video color features, plus a latent decoder to preserve visual fidelity. The authors report robust decoding with 512-bit capacity even under distortions.
Sanitizer is a federated learning framework that combines privacy protection with Byzantine robustness while avoiding heavy cryptographic techniques. Its core is two new protocols, high-dimensional boolean summation and weighted boolean majority vote, built on a Byzantine-robust defense for sign-based FL. On real-world datasets, the authors report 19 to 23 times less runtime than the state of the art, with accuracy matching plaintext and superior robustness against classic attacks.