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
SymSyz is a hybrid kernel fuzzing framework that pairs kernel fuzzing with symbolic execution to reach deep kernel bugs. Its three techniques are Snapshot-Driven Environment Synchronization, Origin-Scoped Symbolic Execution and Priority-Guided Constraint Solving. Against state-of-the-art kernel fuzzers, it improves coverage by up to 27%, reaches equivalent coverage with a speedup of up to 2.2x, and delivers a symbolic tracking speedup of 5.6x, finding 13 previously unknown bugs on the latest Linux kernel.
The paper introduces Models-to-Model Learning (M2ML), a paradigm that builds a unified model directly from existing encrypted traffic analysis (ETA) models without manually labeled data. It uses a large language model with Feature-Negative Entropy Injection and Consensus Credibility-Weighted Arbitration to align heterogeneous feature spaces and resolve conflicts among teacher models. On three public encrypted traffic datasets, the framework reaches accuracy above the best teacher model.
This paper proposes a privacy-preserving deduplication and data integrity auditing scheme for compressed cloud storage. It replaces BLS-based authenticators with a ciphertext built from lightweight algebraic operations, and uses randomized blinding so a third-party auditor cannot infer file ownership. Experiments report up to a 2250-fold reduction in initial upload computation, roughly 1480-fold faster proof generation, and up to 52% lower storage cost.
The authors propose MsaaDI, a federated learning framework for identifying IoT devices under heterogeneous data. It pairs a server-side Multi-Scale Adaptive Aggregation (MSAA) scheme with a client-side fingerprinting pipeline that produces compact grayscale-image representations. On the UNSW and Aalto IoT datasets, using only 20% of training data under extreme label-skewed Non-IID conditions, it reaches 94.35% and 87.72% accuracy, up to 14.33% absolute above the baseline DA-PFL.
IDTD-RL is a risk-sensitive reinforcement learning framework for anti-spoofing defense in OFDM-based UAV communications, using a conditional value-at-risk objective to balance performance and tail risk. Once spoofing is detected, it switches to a dual-track transmission scheme with phantom rate control. In 3GPP TR 36.777 UMi-AV simulations at 0 dB interference-to-signal ratio, it reached 22.37 Mbps covert throughput, 99.64% SINR compliance, and redirected 89.71% of spoofing energy to a decoy.
This research article presents a lattice-based lightweight mutual authentication protocol for Vehicle-to-Grid (V2G) communication between electric vehicles and charging stations over open wireless channels. The authors state that most existing schemes are vulnerable to post-quantum threats, which motivates the design. The protocol's security is checked with formal analysis, automated verification using the AVISPA tool, informal security analysis, and performance simulations.
The authors propose a template protection framework for multimodal palmprint and palmvein biometrics. Their feature-level fusion network, PalmSynNet, pairs a Multi-Scale Local Feature Extraction module with a Global Feature Fusion module, and a cancellable random projection method called SoftMax-based random projection (SoRP) generates protected templates. Evaluations on the PolyU, TJ, and CUMT databases report superior fusion recognition performance and satisfy irreversibility, revocability, unlinkability, and resistance to various attacks.
Synesthete is a new method for detecting unknown smart contract threats. It fuses multimodal features (texts, graph structures and images extracted from opcodes, sources and bytecode streams) with self-learning, data generation and a domain-adaptive learning mechanism. On known reentrancy, timestamp dependence and transaction-ordering dependence vulnerabilities, its detection accuracy improves by 5.46%–28.78% over the latest classical and AI-based methods.
EDT-ABC is a threshold attribute-based anonymous credential system in which each issuer generates its secret key independently, with no distributed key generation protocol or trusted party, and verifiers can dynamically set thresholds per service. The paper reports that, on a personal computer and a Raspberry Pi, EDT-ABC cuts credential issuance and showing time by at least 67.44% and 28.78% respectively, and reduces token size by over 82.10% compared with existing solutions.
DiEL is a face enhancement method that disentangles identity and pose information and optimizes reconstruction, adversarial, and identity-preserving objectives together. It uses a unified face pose dictionary to keep reconstructed faces pose-consistent. Across six benchmarks (MS1M, LFW, CPLFW, CFP-FF, CFP-FP, AgeDB), it averages a 4.66% improvement over state-of-the-art methods, with the largest gains on CPLFW and CFP-FP.
Almohri's WebTrap paper, published in Computers & Security (Volume 171) in December 2026, addresses algorithmic complexity attacks on web endpoints. The source text provided contains only the publication metadata and author, with no abstract or findings, so the research question, method and results cannot be reported from it.
Researchers propose SCVLM, a Sparsity-Controllable Vision-Language Model for scenario-related video anomaly detection, which learns normality from unlabeled normal data instead of requiring rare anomalous samples. The approach jointly reconstructs multimodal representations and summarizes textual descriptions, using a Sparsity-Controllable Memory Block and fusing two modality-specific reconstruction errors with an LLM-based textual anomaly scoring mechanism. The authors report state-of-the-art detection performance on challenging benchmarks.
MF2DA is a unified end-to-end framework for detecting and attributing AI-generated images, aimed at robustness to real-world distortions such as social media compression. It combines an Edge Pyramid Fusion ResNet (EPF-ResNet) for pixel-level edge artifacts, a frozen CLIP-ViT for semantic generalization, an MLLM-Guided Quality Refinement Module (MQRM), and a Dual-stream Differential Patch Attribution Network (D2PAN) for generator fingerprints. The authors report state-of-the-art detection of both "friendly" and "ex-regulatory" images with strong cross-dataset generalization across multiple benchmarks.
Fed-CBE is a client-side defense for federated learning that removes backdoors through three mechanisms: periodic alternating layer resetting, indiscriminate forgetting via entropy maximization on non-ground-truth classes, and knowledge distillation with historical local models. Evaluated on three benchmark datasets and model architectures, it keeps attack success rates near zero in most settings, including under high malicious-client ratios, while preserving primary task accuracy.
Fix: Client-side defense: apply Fed-CBE, which combines periodic alternating layer resetting, indiscriminate forgetting, and knowledge distillation with historical local models to eliminate backdoors without prior trigger knowledge.
IEEE Xplore (Security & AI Journals)ROBIN++ is a watermarking framework for diffusion models that aims to provide both copyright protection and tamper localization. It injects a frequency-domain copyright watermark during generation and a spatial-domain localization watermark in the latent space, then uses a dual-branch detector to verify both. The source reports that it significantly outperforms existing approaches in copyright recovery and tamper localization accuracy, though it gives no specific numbers in the text provided.
OpenAI published two studies on how enterprises use AI, arguing that use is shifting from assistance to execution through agents. Frontier firms, the top 10% of monthly AI usage, generated 8.3× as many output tokens per active user as typical firms as of June, up from 2.6× in January. Codex produced 64% of combined Codex and ChatGPT output tokens among enterprise customers as of June.
The source is a journal citation for a paper titled "AttackLogGen: Benchmarking LLMs for Generating Attack Logs," published in Digital Threats: Research and Practice, Volume 7, Issue 3, pages 1-25, in September 2026. The text provides no abstract, method, or findings, so the research question and results cannot be stated from it.
This paper proposes PaME (DFL by Partial Message Exchange), a decentralized federated learning algorithm in which neighbor nodes exchange only randomly selected sparse coordinates. The authors state that this cuts communication costs and limits exposure of data-sensitive information, a property they characterize with a reconstruction-risk theory under partial observation. They prove linear-rate convergence in expectation to a stationary point under local Lipschitz continuity of the gradient and a doubly stochastic communication matrix, and report numerical experiments showing better performance than several decentralized learning baselines.