Mask2Keep: Mask-guided information transfer for backdoors resilient to compression-oriented pruning

Area of Science:

• Artificial Intelligence
• Machine Learning Security

Background:

• Deep neural networks (DNNs) are susceptible to backdoor attacks, where malicious behaviors are implanted during training.
• Compression-oriented pruning, used for model efficiency, can inadvertently disrupt existing backdoor defenses.
• The fragility of current backdoor attacks under pruning necessitates new defense strategies.

Purpose of the Study:

• To investigate the impact of compression-oriented pruning on the effectiveness of deep neural network backdoor attacks.
• To propose a novel attack-enhancement framework, Mask2Keep (M2K), to improve backdoor robustness against pruning.

Main Methods:

• Developed Mask2Keep (M2K), a three-stage training framework: base backdoor model training with pruning-aware masking, targeted regularization for information redistribution, and knowledge distillation with parameter freezing.
• Evaluated M2K across various attack types, datasets, and pruning strategies to assess its effectiveness.

Main Results:

• Demonstrated that many existing backdoor attacks are fragile and ineffective under compression-oriented pruning.
• Showcased M2K's ability to maintain high attack success rates and competitive clean accuracy even after aggressive model compression.
• Confirmed M2K's robustness against compression-oriented pruning techniques.

Conclusions:

• Compression-oriented pruning poses a significant challenge to existing DNN backdoor attack defenses.
• Mask2Keep (M2K) offers a robust solution for enhancing backdoor attack resilience against model compression, preserving both attack efficacy and model utility.