Hybrid backdoor attacks for deep code models.

Area of Science:

• Artificial Intelligence
• Software Security
• Machine Learning

Background:

• Deep code models are susceptible to backdoor attacks.
• Existing single-trigger attacks lack stealth and are vulnerable to defenses.

Purpose of the Study:

• To propose a novel hybrid backdoor attack method for deep code models.
• To enhance attack effectiveness and stealth by combining multiple trigger types.

Main Methods:

• A hybrid attack combining function signature features and dead code insertion as triggers.
• Leveraging complementary trigger effects for synergistic enhancement.
• Experimental evaluation with minimal poisoning rates.

Main Results:

• The hybrid method achieves high attack success rates (ASR), significantly outperforming single-trigger methods.
• The approach effectively evades spectral feature-based defenses.
• Minimal impact on normal model functionality is maintained.

Conclusions:

• Sophisticated hybrid backdoor attacks pose a significant threat to deep code models.
• Specialized defense mechanisms are urgently needed to counter these advanced threats.