IBD: An Interpretable Backdoor-Detection Method via Multivariate Interactions.

Area of Science:

• Artificial Intelligence
• Machine Learning Security
• Deep Neural Networks

Background:

• Deep neural networks (DNNs) are susceptible to backdoor attacks, where models are compromised during training.
• Current backdoor defense methods lack theoretical grounding and interpretability, often failing against new attack vectors.

Purpose of the Study:

• To propose an interpretable backdoor detection method (IBD) for DNNs.
• To provide a theoretically sound approach for identifying backdoor attacks and poisoned examples.

Main Methods:

• Utilized information theory to analyze multivariate feature interactions, revealing backdoor mechanisms.
• Developed an interpretable theorem to guide backdoor and poisoned example detection.
• IBD detects threats without prior knowledge of specific attack methods.

Main Results:

• IBD demonstrated a significant 78% average increase in detection accuracy compared to existing methods.
• Achieved an order-of-magnitude reduction in detection time cost.
• Effectively detected backdoor models and poisoned examples across widely used datasets and models.

Conclusions:

• IBD offers a robust, interpretable, and theoretically grounded solution for backdoor defense in DNNs.
• The method enhances the security of machine learning models against sophisticated attacks.
• IBD provides a practical tool for defenders to identify and mitigate threats.