Threat Detection in NoC-Based Manycores Using Lightweight Machine-Learning Models

NoC-based manycores are vulnerable to security threats, including subtle and hard-to-detect Hardware Trojans (HTs). Traditional detection methods, such as static latency analysis or hardware-based approaches, either miss subtle anomalies or incur excessive overhead. This paper proposes a lightweight, non-invasive machine learning (ML) approach using XGBoost regression to detect anomalies in NoC traffic. The method employs clock-cycle simulations of real applications to generate training datasets that capture temporal signatures. The ML flow comprises dataset generation, model training, inference, and anomaly detection. Results show high detection rates, recall, and precision across benchmarks, corroborating the method’s effectiveness in identifying HT-induced anomalies with minimal false positives.