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 record for a Computers & Security article by Zexiao Zou, Zhiqiang Wang, Baoxu Liu, Yuyang Han and Yan Zhang, available online 10 June 2026. It is titled "Hiding the trees in the forest: Building network covert channels with hash-based covert carrier filtering." The provided text contains no abstract, method details or findings.
PAAS is a policy-adaptive traceable anonymity scheme for cross-domain networks, designed to meet FATF Travel Rule obligations for virtual asset service providers (VASPs) without giving up user privacy. It maps trace tiers to a hierarchical, role-weighted t-of-n threshold structure and separates pseudonym linkage from real-world identity disclosure, while bounding revocation-bypass risk within a latency window Δ. The authors report service-phase latency of 0.22 ms or 6.12 ms and throughput of about 49.5 TPS under 100 concurrent users, outperforming the evaluated baselines.
Researchers examine a two-sided threat to power grids with renewable distributed energy resources (DERs), called ToLaR, which jointly exploits load-side and DER-side dynamics beyond conventional MadIoT. Tests on six benchmark systems and a physical microgrid testbed show ToLaR produces comparable cascading impacts at only 10% of MadIoT's attack cost. By manipulating the fast frequency response of energy storage systems, the transient grid frequency can be driven down to 53 Hz, triggering severe instability.
The source is a bibliographic record for a September 2026 article in the Journal of Information Security and Applications, Volume 101, by Xinxing Zhao, Kar Wai Fok and Vrizlynn L.L. Thing. The title indicates a threshold-free network anomaly detection method based on comparative reconstruction error learning with parallel GANs. The source text does not state the method's results or numbers.
Anna Cartwright and Edward Cartwright examine whether employee cybersecurity training improves knowledge of best practice for secure passwords and the use of public Wi-Fi. The source text provided contains only the publication date, the journal (Computers & Security) and the author names, so the research method and findings are not stated.
CECO-LAD is a cloud-edge framework for unsupervised log anomaly detection that combines a cloud-side ensemble of expectation-maximization-enhanced anomaly transformers (BAT) with a quantized edge model (Q-BAT). A Mahalanobis distance-based routing policy forwards only uncertain samples to the cloud and keeps confident cases at the edge. The authors report F1-scores of 99.98% on HDFS, 99.95% on OpenStack and 99.99% on BGL while reducing edge resource consumption.
This paper models false data injection (FDI) attacks on networked control systems as a non-cooperative game between legitimate transmitters and a malicious adversary, using the age of incorrect information (AoII) metric. The authors analytically show that a Nash equilibrium exists and is unique, and derive conditions for equilibrium resource allocation. Numerical results, covering drift rates, FDI and transmission costs, and resource constraints, are presented alongside examples from secure healthcare and automotive control.
Researchers introduce MFA-DPRF, a multi-factor authentication scheme that combines attribute-based fine-grained access control with a dynamic password recovery method. Recovery relies on secret questions and secret sharing, letting users restore passwords locally and update secret values. The authors reduce its security to the computational Diffie-Hellman problem under the Random Oracle Model, analyze it in the universally composable framework, and report that performance analysis shows it is superior to state-of-the-art works in security and efficiency.
The source provides only bibliographic metadata for a paper titled "An essential secret image sharing scheme with certification based on the Chinese Remainder Theorem and polynomials." It lists a September 2026 publication in the Journal of Information Security and Applications, Volume 101, by Jianhong Li and Xingbin Liu, but gives no method details or findings.
TAPGuard is a semantics-enhanced framework for modeling rule linkages and detecting cascading threats in Trigger-Action Programming rules for smart home devices. It uses large language models to extract structured semantic elements from natural language rule descriptions, then applies a dual-relation context encoder with node-level and semantic-level attention to reason across multi-hop dependencies. Evaluated on a real-world smart home dataset, it significantly outperforms state-of-the-art graph-based baselines.
Researchers show that black-box targeted attacks on Large Vision-Language Models can be made more precise by working through the projector, a semantic bridge between vision and language. They propose Intermediate Projector Guided Attack (IPGA), which aligns Q-Former query outputs with a target and transfers across models, and IPGA with Residual Query Alignment (IPGA-R), which also preserves non-target content for fine-grained edits. The authors report that IPGA beats baselines on global attacks, IPGA-R wins on fine-grained attacks, and the method transfers to Google Gemini and OpenAI GPT.
SOOM is a schedule-search-based operator obfuscation method that defends compiled DNN models against model extraction attacks on standard CPU and GPU backends. Built on TVM, it uses a security-aware learned cost model based on XGBoost gradient boosted trees to balance security and performance. Tests over 105 operator configurations and more than 30,000 tensor computation test cases raised the operator inference failure rate against state-of-the-art extraction attacks to as high as 89%, with performance gains of up to approximately 25.4% in selected cases.
Fix: SOOM: a schedule-search-based operator obfuscation method built on TVM, using a security-aware learned cost model based on XGBoost gradient boosted trees to generate obfuscated executable code for deep learning operators.
IEEE Xplore (Security & AI Journals)This paper shows that existing model fingerprinting techniques for open-source model IP are vulnerable to false claim attacks, where adversaries assert ownership of independent third-party models. The authors attribute this to untargeted methods that compare arbitrary sample outputs, and propose FIT-Print, a targeted paradigm with two black-box methods, FIT-ModelDiff and FIT-LIME. Reported results include a 100% defense success rate against false claims, a 0.0% false alarm rate on independent models, and a 100% ownership verification rate across diverse model reuse techniques.
Fix: FIT-Print: a targeted fingerprinting paradigm that uses optimization to turn the fingerprint into a verifiable, targeted signature, implemented via the black-box methods FIT-ModelDiff (bit-wise, using output distances) and FIT-LIME (list-wise, using feature attributions).
IEEE Xplore (Security & AI Journals)Researchers propose DANP (Dual Attention-Guided Noise Perturbation), an immunization method that adds imperceptible perturbations to text-to-image diffusion models to resist malicious edits. DANP works across multiple timesteps, reducing cross-attention in text-relevant regions while increasing it in irrelevant ones, and maximizes the discrepancy between injected noise and predicted noise. The authors report state-of-the-art performance against malicious edits.
Fix: The source states that the authors propose countermeasures from the perspectives of the power grid, the network system, and attack detection, but does not give their specific details.
IEEE Xplore (Security & AI Journals)The paper proposes MU-MIA, a membership inference attack that uses machine unlearning to gradually reduce a model's memorization of specific samples and records a forgetting trajectory for each one. A BiLSTM-based binary classifier with attention then separates member from non-member samples using these trajectories. The unlearning step runs in a zero-shot setting that needs no real data, and the authors report that the method outperforms existing baseline attacks across different datasets and model architectures.
Fix: The source states that the analysis yields actionable guidelines for enhancing sensor security through parameter tuning and resource allocation.
IEEE Xplore (Security & AI Journals)A pre-registered randomized trial in Sierra Leone tested Google's Guided Learning mode in Gemini with secondary-school math students over eight weeks. Students gained +0.258 standard deviations in math scores versus the control group, roughly 1.2 to 1.7 years of typical progress, and 91.4% of analyzed conversations (over 113,000 interactions) focused on building conceptual understanding, with Gemini posing scaffolding questions in 76% of messages and giving direct solutions in 2%.