Hybrid backdoor attacks for deep code models

by Tongcheng Geng

Deep code models face security vulnerabilities through backdoor attacks. Previous approaches have primarily relied on single-trigger mechanisms, resulting in limited stealth and vulnerability to defense strategies. This paper proposes a novel hybrid backdoor attack method that combines function signature features and dead code insertion as triggers. Our approach leverages the complementary nature of these triggers, creating a synergistic effect that enhances attack effectiveness while maintaining stealth. Experimentally, with minimal poisoning rates, our hybrid method achieves high attack success rates(ASR), significantly outperforming single-trigger methods. The hybrid approach effectively evades spectral feature-based defenses while maintaining minimal impact on normal functionality. Our findings highlight the urgent need for specialized defense mechanisms against sophisticated hybrid backdoor attacks in deep code models.