About this document
AnyAttack: Self-Supervised Adversarial Attacks on VLMs by ruijiaxiao262 is a document available to read on EtoBox.
The document introduces AnyAttack, a self-supervised framework designed to create large-scale adversarial attacks on Vision-Language Models (VLMs) without the need for specific target labels. By utilizing a pre-training and fine-tuning approach on the LAION-400M dataset, AnyAttack allows for the transformation of any image into an adversarial vector targeting any output across various VLMs. The effectiveness of AnyAttack is validated through extensive experiments on multiple open-source and commercial VLMs,
- Author
- ruijiaxiao262
- Language
- EN