A framework for hardware trojan detection based on contrastive learning.

Area of Science:

• Hardware security
• Integrated circuit design
• Computer engineering

Background:

• Hardware Trojans (HT) pose a significant threat to the semiconductor industry across design, manufacturing, and deployment.
• Side-channel analysis, particularly using power consumption, is a key non-contact method for detecting HT due to its efficiency and accuracy.

Purpose of the Study:

• To propose a novel framework for Hardware Trojan detection using contrastive learning.
• To address the challenges of unsupervised or weakly supervised detection scenarios.
• To improve the generalization capabilities of HT detection models.

Main Methods:

• A framework for HT detection based on contrastive learning utilizing power consumption information.
• Data augmentation techniques, including a one-dimensional discrete chaotic mapping, to enhance model generalization.
• Learning model representations by comparing sample similarities and differences, reducing reliance on labeled data.
• Utilizing a backbone network to categorize side-channel information for efficient HT detection.

Main Results:

• The proposed contrastive learning framework demonstrates superior generalization capabilities for HT detection.
• Models trained on smaller Trojan datasets showed significant detection advantages (up to 44%) on larger Trojans.
• Models trained on larger Trojan datasets also showed advantages (up to 10%) on smaller Trojans.
• The framework performs effectively in imbalanced and noisy data environments.

Conclusions:

• The contrastive learning framework is highly effective for detecting unknown Hardware Trojans in unsupervised or weakly supervised scenarios.
• The method offers improved detection efficiency and generalization compared to traditional approaches.
• This approach advances hardware security by providing a robust solution for identifying malicious implants.