This paper presents a novel adversarial defense framework that strategically exploits the non-transferability of adversarial attacks across multi-modal foundation models. While Contrastive Language–Image Pre-training (CLIP) models demonstrate remarkable zero-shot capabilities, they remain vulnerable to adversarial samples. Adversarial fine-tuning is widely adopted as a standard defense, yet the resulting robustness against sophisticated white-box attacks is often insufficient. To address this limitation, we aim to boost the robustness of an adversarially fine-tuned model by utilizing a pre-trained auxiliary model to leverage attack non-transferability. Specifically, we construct a common embedding space and introduce a detection scheme that identifies the attack target based on feature distances. By adaptively switching the prediction output, we effectively mitigate attacks. Experimental results demonstrate that our approach outperforms state-of-the-art adversarial fine-tuning methods in terms of adversarial robustness.
Toishi et al. (Fri,) studied this question.