BadCM: Invisible Backdoor Attack Against Cross-Modal Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning Security
• Cross-Modal Learning

Background:

• Unimodal backdoor attacks are well-studied, but cross-modal attacks lack generalization and stealth.
• Existing cross-modal attacks often inherit from unimodal visual attacks, limiting their effectiveness.
• Imperceptible trigger samples are crucial for practical real-world attacks.

Purpose of the Study:

• To propose a generalized invisible backdoor framework for cross-modal learning.
• To address the limitations of existing cross-modal backdoor attacks in terms of generalization and stealth.
• To develop a unified framework capable of diverse cross-modal attack scenarios.

Main Methods:

• Introduced a novel bilateral backdoor approach.
• Developed a cross-modal mining scheme to identify modality-invariant components for trigger injection.
• Conceived modality-specific generators for visual and linguistic data to enhance trigger stealth.
• Adapted the framework for image-text cross-modal models.

Main Results:

• Demonstrated the effectiveness and generalization of the BadCM framework on cross-modal retrieval and Visual Question Answering (VQA).
• Showcased the ability of BadCM to evade existing backdoor defenses.
• Achieved high stealthiness in poisoned samples by hiding explicit triggers in modality-invariant regions.

Conclusions:

• BadCM is the first invisible backdoor method designed for diverse cross-modal attacks within a unified framework.
• The proposed method offers improved generalization and stealth compared to existing approaches.
• BadCM presents a significant advancement in understanding and mitigating cross-modal learning vulnerabilities.