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 ACM Computing Surveys article, published July 2026 in Volume 58, Issue 9 (pages 1-37), examines alignment of diffusion models. The source text provided contains only the citation details and no article content, so no methods or findings can be described.
The source is a 2026 journal issue listing, Digital Threats: Research and Practice, Volume 7, Issue 1, pages 1-31. It is cited with the title "Adaptive Real-Time Financial Fraud Detection with Explainable AI Tools," but the text provided contains no abstract, method, or findings.
The paper presents Fed-Adapt, a federated learning framework for adaptive topology reconfiguration against multi-rate DDoS and database flooding attacks. It argues that existing federated approaches to network security suffer from slow convergence, vulnerability to Byzantine participants, and a lack of automated mitigation that turns threat intelligence into network-level countermeasures.
This review surveys recent large language model (LLM) approaches to malware code analysis, covering detection, generation, real-time monitoring, reverse engineering and family analysis. It maps the research landscape, highlights key challenges and emerging innovations, and introduces notable datasets and specialized LLM models. The authors also emphasize the role of static analysis in malware detection.
Researchers propose FIPAPNet, a few-shot learning network for classifying malware from dynamic analysis. It converts system call sequences into feature images, combining Autoencoder pre-training, a Siamese Neural Network, a Multi-prototype Network and DBSCAN clustering. The authors report 93% accuracy on 5-way 5-shot tasks, which they say surpasses existing few-shot dynamic analysis methods.
The paper proposes a multi-modal malware classification approach that combines hierarchical consistency with saliency-constrained adversarial training. It is motivated by the growing volume of malware, with over 1.2 billion samples by 2024 and more than 450,000 new detections per day, and by the weaknesses of signature-based, behavior-based and heuristic detection, including sandbox evasion and high false positives. The excerpt covers only the introduction and does not report the method's results.
This systematic review examines the roles humans play in addressing bias in AI, along with the solutions and methodological approaches proposed for it. The source text provided is only a citation line (ACM Computing Surveys, Volume 58, Issue 9, pages 1-37, July 2026) and contains no article content describing findings or methods.
A survey published in ACM Computing Surveys (Volume 58, Issue 9, pages 1-35, July 2026) reviews methods for detecting training data in large language models. The source text provided contains only the bibliographic citation, with no article content, findings or methodology described.
This ACM Computing Surveys paper, published in Volume 58, Issue 10 (July 2026), examines fairness in deepfake detection across ethnicity, gender and age. The source text provided contains only bibliographic details and no findings, method or results.
AGentVLM is an access control policy generation and verification framework built on small, open-source language models, designed for on-premise deployment to preserve data confidentiality. The authors report an average F1 score of 90.6 % for identifying natural language access control policies from high-level requirements and a policy generation F1 score of 80.6 %. The framework also extracts complex elements such as purposes and conditions and provides feedback to help administrators refine policies before deployment.
The paper proposes AMF-CFL, a robust aggregation algorithm for federated learning under non-i.i.d. client data distributions. It filters malicious client updates in two steps, first with multi-k-means clustering and then with z-score-based statistical analysis. The authors report that it preserves global model integrity against four untargeted and two targeted attack types.
XAIDroid, a framework from the authors, uses graph neural networks and graph attention mechanisms (GAM and GATv2) over API call graphs to localize malicious code in Android malware at class and method levels. The authors report recall of 97.27% and F1-score of 95.30% at class level, and 96.02% recall with 92.68% F1-score at method level, evaluated on synthetic and real-world malware datasets.
The paper proposes VFEFL, a privacy-preserving federated learning scheme that encrypts local models with a verifiable functional encryption scheme. It aims to protect local model privacy and defend against malicious clients who upload poisoned models or wrong ciphertexts or decryption keys, within the basic one-server, multiple-client architecture. The authors report that prototype experiments show it resists targeted and untargeted poisoning attacks while preserving model privacy.
Vuln2Action is an LLM-based framework for generating vulnerability reproduction steps and mapping exploits. The authors note that more than 40,000 CVEs were published in 2024 and that only about 0.4% of NVD vulnerabilities from 2010 to May 2025 have publicly listed exploits in ExploitDB. The introduction also reports that preliminary tests of several open-source and instruction-tuned models showed consistent refusals when prompts requested exploit code or attack instructions.