MFBD : Model-free backdoor defense based on vision-language pre-trained models.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Cybersecurity

Background:

• Deep neural networks (DNNs) are vulnerable to backdoor attacks, where malicious triggers embedded in images cause misclassification.
• Existing defenses often require additional data or models, limiting their effectiveness against diverse attack strategies.

Purpose of the Study:

• To propose an effective model-free backdoor defense method (MFBD) for DNNs.
• To leverage the robustness of vision-language models against backdoor triggers.
• To enhance the security of DNNs against sophisticated poisoning attacks.

Main Methods:

• Utilized vision-language models (BLIP-2, CLIP) to generate semantic descriptions of images and their saliency-masked counterparts.
• Employed Sentence-BERT (SBERT) to embed image descriptions and labels into a semantic space.
• Computed cosine similarity between semantic representations to identify poisoned images based on low similarity scores.

Main Results:

• Vision-language models demonstrated significant resistance to backdoor triggers in poisoned images.
• The proposed MFBD method effectively mitigated various complex backdoor threats across multiple datasets.
• MFBD maintained high performance on clean data while successfully identifying and neutralizing poisoned samples.

Conclusions:

• MFBD offers a robust and effective model-free approach to defending DNNs against backdoor attacks.
• Leveraging semantic consistency via dual vision-language descriptors presents a promising direction for future cybersecurity research in AI.
• The method shows potential for practical application in securing DNNs in real-world scenarios.