Deep Graph Learning for Circuit Deobfuscation.

Area of Science:

• Computer Engineering
• Cybersecurity
• Artificial Intelligence

Background:

• Circuit obfuscation protects integrated circuit (IC) intellectual property (IP) from reverse engineering.
• Satisfiability (SAT)-checking attacks can deobfuscate protected circuits, but runtime varies significantly.
• Accurate deobfuscation runtime prediction is crucial for optimizing IC design defenses.

Purpose of the Study:

• To develop an efficient framework for predicting deobfuscation runtime of obfuscated ICs.
• To address challenges posed by graph-structured circuits, varying topologies, and efficiency requirements.
• To aid IC designers in optimizing defense strategies against reverse engineering.

Main Methods:

• Utilized graph deep learning techniques to predict deobfuscation runtime.
• Characterized SAT problem complexity using a conjunctive normal form (CNF) bipartite graph.
• Developed the Deep Survival Analysis with Graph (DSAG) framework, integrating energy-based layers and censored regression.

Main Results:

• The proposed framework accurately predicts deobfuscation runtime for obfuscated circuits.
• DSAG effectively extracts determinant features, improving upon standard regression models.
• Demonstrated effectiveness and efficiency through extensive experiments on benchmark circuits.

Conclusions:

• The DSAG framework offers a robust solution for predicting deobfuscation runtime in IC design.
• This approach enhances the optimization of circuit obfuscation defenses.
• Enables faster and more accurate security assessments for integrated circuits.