ATAC: Augmentation-Based Test-Time Adversarial Correction for CLIP

Su Linxiang, András Balogh

Keywords: CLIP, Test-Time Defense, Adversarial Robustness, Vision-Language Modelsarial robustness

TL;DR: ATAC is a test-time defense that uses data augmentations to estimate a recovery direction and correct adversarial embeddings in CLIP’s feature space, improving robustness without retraining and with low computational cost.

Abstract: Despite its remarkable success in zero-shot image-text matching, CLIP remains highly vulnerable to adversarial perturbations on images. As adversarial fine-tuning is prohibitively costly, recent works explore various test-time defense strategies; however, these approaches still exhibit limited robustness. In this work, we revisit this problem and propose a simple yet effective strategy: Augmentation-based Test-time Adversarial Correction (ATAC). Our method operates directly in the embedding space of CLIP, calculating augmentation-induced drift vectors to infer a semantic recovery direction and correcting the embedding based on the angular consistency of these latent drifts. Across a wide range of benchmarks, ATAC consistently achieves remarkably high robustness, surpassing that of previous state-of-the-art methods by nearly 50\% on average, all while requiring minimal computational overhead. Furthermore, ATAC retains state-of-the-art robustness in unconventional and extreme settings and even achieves nontrivial robustness against adaptive attacks. Our results demonstrate that ATAC is an efficient method in a novel paradigm for test-time adversarial defenses in the embedding space of CLIP. Code is available at https://github.com/kylin0421/ATAC

Cps Compliance Confirmation: true

Submission Number: 2