ATAC boosts CLIP robustness to adversarial image changes by correcting embeddings via augmentation-induced drift vectors and angular consistency checks, beating prior test-time methods by nearly 50 percent on average with low overhead.
Test-time prompt tuning for zero-shot generalization in vision-language models.Advances in Neural Information Processing Systems, 35:14274–14289, 2022
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2025 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
ATAC: Augmentation-Based Test-Time Adversarial Correction for CLIP
ATAC boosts CLIP robustness to adversarial image changes by correcting embeddings via augmentation-induced drift vectors and angular consistency checks, beating prior test-time methods by nearly 50 percent on average with low overhead.