Versatile Backdoor Attack With Visible, Semantic, Sample-Specific and Compatible Triggers.

Area of Science:

• Artificial Intelligence
• Machine Learning Security
• Computer Vision

Background:

• Deep neural networks (DNNs) are vulnerable to backdoor attacks, where specific triggers manipulate behavior without impacting benign performance.
• Implementing physical backdoor attacks is challenging due to manual trigger selection, time constraints, and sensitivity to real-world visual distortions.

Purpose of the Study:

• To address the limitations of current physical backdoor attacks.
• To introduce a novel trigger and an automated pipeline for effective, stealthy, and robust physical backdoor attacks.

Main Methods:

• Defined a Visible, Semantic, Sample-specific, and Compatible (VSSC) trigger for physical deployment.
• Developed an automated pipeline with modules for trigger selection (using LLMs), insertion (using generative models), and quality assessment (using vision-language models).

Main Results:

• The VSSC trigger demonstrated effectiveness, stealthiness, and robustness against visual distortions.
• The automated pipeline successfully integrated triggers into images, ensuring natural and successful insertion.
• Validated strong practicality in physical scenarios, transforming backdoor attacks into a systematic threat.

Conclusions:

• The VSSC trigger and automated pipeline offer a practical solution for physical backdoor attacks.
• This work advances the field by making physical backdoor attacks more systematic and realistic.
• The proposed approach can inspire future research in designing practical triggers for backdoor attacks.