InfoResearchPeer-reviewedLLM-specific
Adversarial Attacks on Vision-Language Multimodal Systems: A Survey
- Published
- Record updated
Summary
This survey reviews adversarial attacks on vision-language multimodal systems (VLMS), covering pretrained models, large language models, attack methods, defense strategies, benchmark datasets, downstream tasks, and evaluation metrics. It analyzes the adversarial vulnerabilities of VLMS and examines task-specific attacks and attacks on different models, including their principles, implementations, and effects. The authors also suggest future research directions such as improving attack universality and compliance considerations.
Topics
Related items
- InfoOut of thin air: An adversarial attack method against human pose estimation networkSimilar attack · Elsevier Security Journals
- LowGenerative Red Teaming for Robust Network Intrusion Detection: A Constrained Adversarial ApproachSimilar attack · OpenAlex (peer-reviewed AI security)
- MediumGHSA-25rm-9wvm-m38v: falcon-rust: Discrete-Gaussian sampler runs at 2^-33 precision, far below Falcon's requirement (reference KAT vectors no longer pass)Similar attack · GitHub Advisory Database
- InfoAdversarial Malware Detection Using Feature Contribution Difference Analysis with Kernel SHAPSimilar attack · OpenAlex (peer-reviewed AI security)
- InfoDensity Feedback Control for Transferability of Generative Sparse Adversarial AttacksSimilar attack · Arxiv (cs.RO + cs.CV security)