Adversarial Purification by Consistency-Aware Latent Space Optimization on Data Manifolds

Summary

This paper proposes Consistency Model-based Adversarial Purification (CMAP), which removes adversarial perturbations from inputs to deep neural networks by optimizing vectors in the latent space of a pre-trained consistency model. The method combines a perceptual consistency restoration mechanism, a latent distribution consistency constraint, and an ensemble-based latent vector consistency prediction scheme. Experiments on CIFAR-10 and ImageNet-100 report significantly improved robustness against strong adversarial attacks while preserving high natural accuracy.

Related items

• MediumRequest, Aggregate, Bypass: How Attackers Can Evade LLM Safety ClassifiersSimilar attack · CrowdStrike Blog
• LowStealthy Physical Adversarial Attacks on Speaker Recognition via Near-Ultrasonic PerturbationsSimilar attack · IEEE Xplore (Security & AI Journals)
• InfoLLMs Cannot Reliably Judge (Yet?): A Comprehensive Assessment on the Robustness of LLM-as-a-JudgeSimilar attack · IEEE Xplore (Security & AI Journals)
• LowWhen the Bee Stings: CyberCom’s AI VulnerabilitySimilar attack · AIS eLibrary (Journal of AIS, CAIS, etc.)
• MediumA Decision Model Breaks Like Any Other Language Model: A First Look at JevSimilar attack · Check Point Research