Abstract

Concerns about hardware security are raised by the increasing dependence on third-party Semiconductor Intellectual Property in system-on-chip design, especially during physical design verification. Traditional rule-based verification methods, such as Design Rule Checking (DRC) and Layout vs. Schematic (LVS) checking, together with side-channel analysis, indicated apparent deficiencies in dealing with new forms of threat. The impossibility of distinguishing dependable from malicious insertions in ICs makes it hard to prevent such dangers as Hardware Trojans (HTs); side-channel vulnerabilities remain everywhere, and modifications at various stages of the manufacturing process can be hard to detect. This research addresses these security challenges by defining a theoretical AI driven framework for secure physical design verification that couples Graph Neural Network models (GNNs) and probabilistic modeling with constraints optimized to maximize IC security. Also, we can model mathematical foundations for the secure routing as a constrained path finding problem for all myths addressed above concerning these different methods-moves are optimized to avoid sources of security problems. These problems might include crosstalk-induced leakage and electromagnetic side-channel threats. As an alternative to experience-based anomaly detection proposed in earlier work, a theoretical softmax based anomaly classification framework is put forward here to model HT insertion probabilities, gathering acceptable anomalies at various levels of circuit design from RTL level to Gate-level as necessary. This theoretical framework provides a conceptual methodology for scalable, automated, and robust security verification in modern ICs through graph-based learning, and constrained optimization methods. It lays a foundation to advance secure semiconductor designs further using AI-driven techniques without recourse to benchmarks or empirical validations.

Keywords: Graph Neural Networks (GNNs), Reinforcement Learning (RL), Softmaxbased anomaly detection, Karush-Kuhn-Tucker (KKT), Deep learning, Lagrange multipliers