Critical Path-Based Backdoor Detection for Deep Neural Networks.

Area of Science:

• Artificial Intelligence
• Machine Learning Security
• Deep Neural Networks

Background:

• Backdoor attacks pose significant threats to deep neural networks (DNNs).
• Existing detection methods often rely on distribution analysis, limiting their generalization across different DNN models.
• A need exists for more robust and interpretable backdoor detection techniques.

Purpose of the Study:

• To propose a novel critical-path-based backdoor detector (CPBD) for deep neural networks.
• To enhance backdoor detection by leveraging DNN interpretability.
• To develop a method capable of generalizing across various DNN architectures and trigger configurations.

Main Methods:

• Developed a critical-path-based backdoor detector (CPBD) focusing on DNN interpretability.
• Simplified neurons and identified key nodes to form critical paths within DNNs.
• Analyzed critical paths corresponding to different classes to detect backdoor characteristics and locate malicious triggers via trigger propagation paths.

Main Results:

• The proposed CPBD method efficiently discovers characteristics distinguishing backdoor attacks.
• The detector demonstrates effectiveness across multiple DNN models and varying trigger sizes.
• CPBD successfully locates neurons associated with malicious triggers, forming trigger propagation paths.

Conclusions:

• CPBD offers an effective and generalizable approach to detecting backdoor attacks in DNNs.
• The method's reliance on interpretability provides a systematic way to identify and understand backdoor mechanisms.
• CPBD advances the field of AI security by offering a robust defense against sophisticated backdoor threats.