Dual-pathway mask ranking guided selective fine-tuning for backdoor purification

Area of Science:

• Artificial Intelligence
• Machine Learning Security
• Computer Vision

Background:

• Backdoor attacks pose a critical threat to AI security by compromising model integrity during training.
• Existing backdoor purification methods like pruning and fine-tuning have limitations, including degraded accuracy and residual backdoor effects.
• Effective backdoor purification requires robust separation of clean and poisoned data.

Purpose of the Study:

• To propose a novel dual-pathway mask ranking guided selective fine-tuning method for backdoor purification.
• To develop an effective strategy for clean-poisoned data separation.
• To mitigate backdoor effects without compromising clean accuracy.

Main Methods:

• An erasure-based intervention strategy was developed, prioritizing spatial reconstruction of backdoor triggers for effective erasure.
• A forget-then-recover mechanism using soft-valued masks was established to characterize neuron contamination.
• A training-free dual-pathway mask ranking module categorizes neurons to guide selective fine-tuning policies (relearning, unlearning, nolearning).

Main Results:

• The proposed method demonstrated superior performance against seven baseline competitors across four benchmark datasets (MNIST-M, SVHN, CIFAR-10, CIFAR-100).
• Achieved the lowest average Attack Success Rate (ASR) between 0.23% and 1.29%.
• Incurred the smallest average Clean Accuracy (CA) degradation between 0.15% and 1.53%.

Conclusions:

• The dual-pathway mask ranking guided selective fine-tuning method effectively purifies backdoor attacks.
• The clean-poisoned separation strategy and selective fine-tuning policy are crucial for successful backdoor mitigation.
• This approach offers a promising solution for enhancing AI security against backdoor threats.