The answer lies within: Detecting Trojans from DNNs' inherent characteristics

Area of Science:

• Artificial Intelligence
• Machine Learning Security

Background:

• Deep neural networks (DNNs) are susceptible to Trojan attacks, where malicious triggers cause them to malfunction.
• Existing Trojan detection methods often rely on benign samples and time-consuming optimization, limiting their practical use.
• The benign sample-free scenario presents a significant challenge for robust Trojan detection.

Purpose of the Study:

• To develop an efficient and generalizable method for detecting Trojans in DNNs without requiring benign samples.
• To address the limitations of traditional trigger reversion techniques in practical applications.

Main Methods:

• Proposed Detecting Trojans from DNNs' inherent characteristics (DTIC), a novel approach leveraging unique features of Trojaned models.
• DTIC utilizes a unified representation space derived from DNN structures and parameters for adaptability across diverse models.
• The method employs random perturbations and the lottery hypothesis for enhanced detection performance.

Main Results:

• DTIC achieves high efficiency, requiring only a single direct inference to detect Trojans.
• Experiments on the IARPA TrajAI benchmark demonstrate DTIC's superior effectiveness and generalizability.
• The proposed method successfully operates in the challenging benign sample-free scenario.

Conclusions:

• DTIC offers a significant advancement in Trojan detection for DNNs, overcoming the limitations of prior methods.
• The approach provides an efficient, effective, and generalizable solution for securing DNNs against Trojan attacks.
• This work paves the way for more reliable DNN security in real-world applications.