Layer Frozen Multi-Net & Latent Space Feature-Concealed Backdoor Samples Detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning Security

Background:

• Backdoor attacks pose a significant threat to machine learning models, especially those with feature-concealed or dynamic triggers.
• Existing detection methods struggle with backdoor samples that overlap with benign data distributions or have fractured distributions.
• These limitations lead to missed detections and high false positives, compromising model integrity.

Purpose of the Study:

• To propose a novel method, Layer Frozen Multi-Net and Multi-Latent Space (LFMN-LS), for robust backdoor sample detection.
• To address the challenges posed by feature-concealed and dynamic-trigger backdoor attacks.
• To improve the accuracy and reduce false positives in backdoor detection systems.

Main Methods:

• Developed LFMN-LS, a method utilizing a Layer Frozen Multi-Net and Multi-Latent Space approach.
• Constructed Trigger-Net and Benign-Net via knowledge refinement to separately capture backdoor and benign sample distributions.
• Introduced Relative Cosine Distance to measure distribution differences across multiple latent spaces, mitigating distributional fractures.

Main Results:

• LFMN-LS demonstrated superior performance compared to existing state-of-the-art backdoor detection methods.
• The method effectively handles backdoor samples with overlapping distributions and fractured patterns.
• LFMN-LS achieved lower false detection rates for benign samples.

Conclusions:

• LFMN-LS offers a more effective and accurate solution for detecting challenging backdoor attacks.
• The integration of layer freezing in knowledge refinement preserves crucial high-order features for better detection.
• This approach enhances the security and reliability of machine learning models against sophisticated threats.