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.
2,038 items
This research paper proposes a personalized differential privacy approach for publishing high-dimensional data, combining random sampling and a pruning privacy tree. The source text is an introduction that motivates the work by arguing that traditional anonymization is weak against inference attacks and that uniform privacy budgets fail to meet heterogeneous privacy needs. It reviews prior personalized differential privacy mechanisms, including Jorgensen et al. and Alaggan et al.
The source text is a biographical profile of Dr. Kamta Nath Mishra, an Associate Professor in the CSE department at B.I.T. Mesra, Ranchi, India. It lists his degrees, including a Ph.D. from B.I.T. Mesra in May 2015 and postdoctoral research at NTNU, Norway, along with his prior positions and research interests in biometric systems, cloud computing and distributed cloud computing.
The paper proposes Video Information Maximization (VIM), a framework for few-shot video action recognition. VIM uses an adaptive spatial-temporal video sampler to select important frames and amplify critical spatial regions, and a spatial-temporal action alignment model to measure inter-video similarity more precisely. A mutual-information-based training objective jointly maximizes intra- and inter-video information, and experiments on public datasets are reported to demonstrate its effectiveness.
This study examines what drives open banking adoption in Taiwan, combining the Technology–Organization–Environment framework with the Analytic Hierarchy Process to build a grounded model. Drawing on 20 experts and 264 potential users, it finds that security concerns consistently act as a critical determinant, while government and senior leadership support are indispensable in the early market.
This paper proposes ANCC, an Anonymous Network Covert Channel that uses Tor's Hidden Service Directories (HSDirs) as intermediate nodes to modulate covert information through the publication and retrieval statuses of hidden services. The authors claim it provides anonymity for the sender, receiver and their communication relationship, with a detection probability below 0.25% against an adversary compromising fifty intermediate nodes. Real-world Tor evaluation reports over 99.6% transmission accuracy and a channel capacity of around 3 Kbps.
EPRU, a reputation-based authentication scheme for vehicular platoons, updates vehicle credit scores through a dynamic aggregation trigger that uses verified real-time feedback. It combines ElGamal and homomorphic encryption to protect vehicle identities and feedback data. The authors report formal security proofs against forgery, replay, and modification attacks, and a reduction of at least 3.05% in computation overhead and 37.5% in communication cost compared with recent state-of-the-art schemes.
This paper introduces MSDT, a blockchain-based data trading protocol that guarantees trading atomicity and data quality without a trusted third party. It uses a state channel-based trading contract to reduce on-chain costs and a staking mechanism with a trading strategy to deal with malicious data owners. The authors report that security and performance analyses show MSDT ensures atomicity and data quality verification while keeping efficiency comparable to existing solutions.
EBFT is a BFT consensus framework for permissioned blockchains that replaces a fixed leader with randomly and non-interactively proposed blocks, combining Nakamoto's longest-chain rule with classical BFT quorum voting. The authors built three protocols (EBFT-Syn, EBFT-PSyn and EBFT-Turbo) atop btcd, with about 920 lines of code, and evaluated them on EC2 with up to 256 nodes. EBFT-Syn and EBFT-PSyn reach latencies of 6 and 1 seconds, while EBFT-Turbo processes up to 3.2k transactions per second at 8 seconds latency.
This paper proposes CP-A$^{2}$2BE, a ciphertext-policy anonymous attribute-based encryption scheme for IIoT data sharing. It adds outsourced decryption through edge nodes deployed in factories, protects attribute privacy for devices and personnel, and introduces a commitment-based verifiable revocation mechanism. Experiments report a constant 26 milliseconds decryption time regardless of the number of attributes.
KubeSec is a framework that analyzes application permission configurations and component code dependencies to find takeover risks in third-party applications (TPAs) running in Kubernetes clusters. The study found 562 insecure RBAC binding patterns and 375 vulnerabilities linked to 134 CVEs, affecting millions of users, with average remediation exceeding 10 months. The findings were reported to the relevant teams, and the community assigned 21 new CVEs.
FreeFL is a cross-silo federated learning protocol that removes the need for a trusted third party and a centralized aggregator. It replaces homomorphic encryption with a lightweight decentralized symmetric encryption scheme with additive homomorphism, built from a decentralized multiparty symmetric encryption (DMSE) scheme and two multiparty computation protocols. The authors report high computation and communication efficiency and significant computational gains over existing HE-based cross-silo FL protocols.
The authors correct an earlier claim that their local information privacy mechanism satisfies (ε,0)-LIP for any ε > 0 and any input distribution. They restrict the valid range of privacy parameters, propose algorithms that expand that range, and revise their experimental results accordingly.
FauForensics is a framework for detecting audio-visual deepfakes that uses facial action units, which quantify facial muscle activity, as forgery-resistant features. It computes frame-wise audio-visual similarities through a fusion module with learnable cross-modal queries, aligning lip-audio relationships to address feature heterogeneity. Tested on four public datasets, it reports state-of-the-art results with a 5.17% average cross-dataset improvement over existing methods.
This position article maps large language model (LLM) capabilities onto the MITRE ATT&CK and NIST Cybersecurity Frameworks. It finds that LLMs show fluency and versatility across cybersecurity tasks but remain fragile in complex, high-stakes settings, due to limited context retention, hallucinations, and reasoning gaps.
The paper presents CIRCUS, a causal intervention framework for improving counterfactual fairness in trained classifiers, built on a new generator called CITGAN. CIRCUS generates counterfactually discriminatory samples via causal intervention, then applies bias correction to their labels and retrains the classifier. On a DNN, the MMD_L and MMD_K values fell by averages of 39.7% and 40.4% versus the second-best result; for ResNet, the reductions were 56.7% and 54.5%.
Fix: The mitigation is the CIRCUS framework itself: generating counterfactually discriminatory samples through causal intervention, applying bias correction preprocessing to their labels, and retraining the classifier.
IEEE Xplore (Security & AI Journals)The authors present a theoretical framework that derives bounds on fairness errors and convergence rates, accounting for how data distributions differ across demographic groups. They validate it on six datasets spanning image, tabular and text data, finding that feature distribution differences significantly affect model fairness, especially across racial categories. They also propose Fairness-Aware Regularization (FAR), a training objective that minimizes inter-group discrepancies in feature centroids and covariances, and report consistent improvements in overall AUC, ES-AUC and subgroup performance.
Fix: Fairness-Aware Regularization (FAR), a training objective that directly minimizes inter-group discrepancies in feature centroids and covariances.
IEEE Xplore (Security & AI Journals)SMInject is a new injection attack framework against pre-trained multimodal models used in cross-modal retrieval. It generates deceptive injections that combine concepts through causal correlation across modalities, and aligns them in the encoded embedding space to boost effectiveness. On representative multimodal models, it reportedly achieves over 14% higher attack success rate and 6% higher Hit@5 than state-of-the-art methods while preserving overall model utility.
Fraud-RLA is a new adversarial attack against credit card fraud detection systems that uses reinforcement learning to evade detection. It aims to maximize the amount stolen while requiring substantially less prior knowledge than competing methods. The authors report that it remains effective on a realistic fraud detection system under the constraints of their threat model.
This paper proposes a triple-branch network for deepfake detection that jointly learns spatial features and frequency features from the original image and from images reconstructed through different frequency channels. The authors add feature decoupling and fusion losses derived from mutual information theory to make the model focus on task-relevant features. The method reportedly achieves state-of-the-art performance across six large-scale benchmark datasets.
This paper proposes HORP, a robust reinforcement learning method that uses the historically optimal policy to guide policy optimization and generates diverse adversarial perturbations. It builds a guidance value function from value gaps and policy distribution divergence, and adds performance-aware correction and controlled uncertainty injection. Experiments report superior natural performance and robustness against various state attacks in most cases.
Fix: The source states the authors provide a correction to the valid range of privacy parameters, propose efficient algorithms to expand the range of valid privacy parameters, and present corrected results. No further fix or configuration change is specified.