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 source is a bibliographic listing for a paper titled "Robust Large-Scale Detection of Living-Off-the-Land Reverse Shells via Data Synthesis" in ACM Transactions on Privacy and Security, Volume 29, Issue 3, pages 1-30, August 2026. The text provides no abstract, method details, or findings.
This paper proposes a quantum-secure, auditable data sharing framework for the Internet of Vehicles. Its core is QRPPRFS, a quantum-resistant puncturable signature algorithm built on a low-noise LPN-based pseudorandom function, combined with blockchain and a dual-commitment proof for anonymity and auditability. With n=256 and q≈2^23, signing and verification each stay below 10 ms and initial secret-key storage drops to 0.22 MB.
OptFuzz is a fuzzing approach for JavaScript JIT compilers that uses optimization path feedback rather than code coverage alone. It approximates optimization paths with an optimization subpath metric, guiding seed preservation and scheduling, and keeps subpaths only when optimizations are triggered at the JIT intermediate representation level. Implemented on four baseline fuzzers and tested on four mainstream JavaScript engines, it found several times more bugs on earlier versions and 52 new bugs on the latest engines, 43 of which were confirmed or fixed.
Researchers propose EGuardSQL, an architecture that provides element-level access control over encrypted SQL tables using Ciphertext-Policy Attribute-Based Encryption (CP-ABE) with hybrid row-level and column-level policies. They add a reusable sub-policy mechanism to cut redundant cryptographic computation, and report significant gains in computational and communication efficiency over conventional CP-ABE approaches, with low storage overhead. The authors formally prove selective IND-CPA security under the decisional q-parallel BDHE assumption and claim resistance to collusion attacks.
The authors propose MCRF, a multi-party digital signature scheme for Industrial IoT environments that adds Cryptographic Reverse Firewalls to resist insider attacks and data exfiltration. MCRF uses a commitment-based mechanism to optimize signing, with an input-side CRF that sanitizes messages and an output-side CRF that re-randomizes signatures without compromising correctness. The source claims strong leakage resistance with minimal computational and communication overhead.
Researchers propose OABE-VPUDA, an outsourced attribute-based encryption scheme for Digital Twins that offloads costly CP-ABE operations to a cloud server. The design adds verifiable outsourced encryption and decryption, dynamic attribute revocation through on-chain expiration tags, and verifiable ciphertext policy updates, all anchored on a consortium blockchain. The authors claim chosen-plaintext security and report efficiency from a prototype implementation.
This paper presents an off-policy safe reinforcement learning approach for nonlinear dynamical systems under input saturation. It formulates safe exploration as a robust control problem using input-to-state safe control barrier functions, enforced through a quadratic programming optimization, with an adaptive epsilon-tuning law that tightens constraints near safety boundaries. Simulations are reported to demonstrate the framework's efficacy.
Video steganography that embeds messages by modifying prediction unit (PU) partition modes usually assumes additive embedding distortion, which fails to capture sample-level distortion drift across frames. The authors propose a modification strategy based on Synchronizing Mode Directions (SMDs) and a frame-level directional model built on Distortion Propagation Chains (DPC) that dynamically corrects the cover cost. The paper reports that the method improves resistance to both intra-frame and inter-frame steganalysis while keeping visual quality and bitrate control.
Advocates of digital financial access promote mobile payment apps (MPAs) as a route to universal access to formal financial services, yet adoption lags in many underserved countries. The authors analyzed over 34,000 online reviews from MPA users, using neural networks and natural language processing to identify affordances and constraints, then tested their relationships with MPA use alongside cross-country cultural values. They found that in traditional value-oriented cultures, the positive effects of affordances weaken and the negative effects of constraints attenuate, while in survival-value-oriented cultures, certain affordances and the credit application complexity constraint have amplified effects.
Researchers propose a video privacy scheme that detects human targets with YOLOv8, selectively encrypts those regions using a hyperchaotic map and a chaos-driven interframe modulation strategy, and restores the background with GLCIC inpainting. Encrypted data is hidden in the restored background through Walsh-Hadamard transform based multi-round embedding. Reported results include an information entropy of 7.9925, a steganographic capacity of 0.75 bpp, and a PSNR above 44.91 dB after embedding.
This paper assesses safety for uncrewed aircraft operating under the U-Space framework, the European counterpart of UTM. The authors identify two metric categories, collision metrics and surveillance performance metrics, and test them experimentally. They vary GPS and IMU failures, UAV speed, update intervals, traffic density and weather to see which metrics track safety risk.
Researchers propose GASE, a style-based generative adversarial framework for medical image segmentation trained on limited single-sequence MR data. The framework adapts to unseen acquisition-level variations through style interpolation and estimates the reliability of predicted masks via confidence learning. Experiments on knee and pelvis MR datasets report high segmentation accuracy, adaptability to unseen variations, and intrinsic interpretability.
Researchers present an integrated detection framework (IDF) based on deep neural networks for blind speech tampering detection, covering copy-move, deletion, and splicing manipulations. The framework combines an encoder-decoder network producing a preliminary localisation mask from Mel spectrogram features, an adaptive localisation refinement algorithm, and a classifier for tampering type. Trained on a new 55,500-pair dataset from Chinese and English corpora, it reports average F-scores of 95.32% (Chinese), 96.02% (English), 88.81% (Spanish) and 95.75% (spoofed samples), with localisation error below 0.1 seconds.
Researchers present AutoMalTool, an automated red teaming framework that generates malicious MCP tools to test LLM-based agents. The authors argue that prior tool poisoning studies stayed at the proof-of-concept stage, leaving automatic red teaming open. Their evaluation reports that AutoMalTool produces malicious MCP tools that steer mainstream LLM-based agents toward unintended behavior while evading current detection mechanisms.
The source is a bibliographic listing for a paper titled "Privacy Against Agnostic Inference Attacks in Vertical Federated Learning" in ACM Transactions on Privacy and Security, Volume 29, Issue 3, pages 1-34, August 2026. The text contains only publication metadata and no abstract, method, or findings.
OpenAI's researchers report that an automated system found limited accidental Chain-of-Thought (CoT) grading during RL training of some released models, including GPT-5.4 Thinking, GPT-5.1 Instant through GPT-5.4 Instant, GPT-5.3 mini, and GPT-5.4 mini, despite the company's policy against it. Their analysis found no clear reduction in CoT monitorability, though they cannot rule out harder-to-measure effects.
Fix: Fixed the affected reward pathways, expanding the automated detection system, and strengthened internal guidance and processes.
OpenAI Alignment Research BlogThis research paper introduces DAAS (Defense via Auto-Augmentation Search) against Model Inversion Attacks, which can recover private training data from model weights or outputs. The authors note that existing defenses offer incomplete protection and that input-level defenses have been limited to simple transformations. DAAS automatically identifies augmentation candidates with strong privacy-utility trade-offs from a large pool, and the authors report superior defense performance across models, datasets and attacks.
Researchers show that balance theory, which signed graph neural networks (SGNNs) use to model positive and negative relationships, creates exploitable weaknesses under closed-box attacks. They propose balance-attack, which targets graph balance degree and is solved with a heuristic for an NP-hard problem, and introduce Balance Augmented-Signed Graph Contrastive Learning (BA-SGCL) to restore robustness by maintaining high balance degree in the latent space. Experiments across multiple SGNN architectures and real-world datasets report both the attack's effectiveness and BA-SGCL's improved resilience.
Fix: BA-SGCL, a contrastive learning framework with balance augmentation, maintains high balance degree in the latent space to improve model resilience against balance-attack.
IEEE Xplore (Security & AI Journals)