Unlearning Backdoor Threats: Enhancing Backdoor Defense in Multimodal Contrastive Learning via Local Token Unlearning

TL;DR

This paper proposes a novel, low-cost defense method for multimodal contrastive learning models against backdoor attacks by identifying suspicious samples and applying localized token unlearning to remove malicious behaviors while maintaining high accuracy.

Contribution

It introduces a token-based localized unlearning approach that effectively mitigates backdoor threats with minimal impact on clean model performance.

Findings

Reduces backdoor attack success rate significantly.

Preserves high clean accuracy after unlearning.

Requires only a small set of poisoned samples for effective defense.

Abstract

Multimodal contrastive learning has emerged as a powerful paradigm for building high-quality features using the complementary strengths of various data modalities. However, the open nature of such systems inadvertently increases the possibility of backdoor attacks. These attacks subtly embed malicious behaviors within the model during training, which can be activated by specific triggers in the inference phase, posing significant security risks. Despite existing countermeasures through fine-tuning that reduce the adverse impacts of such attacks, these defenses often degrade the clean accuracy and necessitate the construction of extensive clean training pairs. In this paper, we explore the possibility of a less-cost defense from the perspective of model unlearning, that is, whether the model can be made to quickly \textbf{u}nlearn \textbf{b}ackdoor \textbf{t}hreats (UBT) by constructing a small set of poisoned samples. Specifically, we strengthen the backdoor shortcuts to discover suspicious samples through overfitting training prioritized by weak similarity samples. Building on the initial identification of suspicious samples, we introduce an innovative token-based localized forgetting training regime. This technique specifically targets the poisoned aspects of the model, applying a focused effort to unlearn the backdoor associations and trying not to damage the integrity of the overall model. Experimental results show that our method not only ensures a minimal success rate for attacks, but also preserves the model's high clean accuracy.