A Machine Learning Approach for Traffic Anomaly Detection in NoC-based Manycores

NoC-based manycores are increasingly vulnerable to traffic anomalies, which can compromise the performance and security of these systems. Early proposals to detect traffic anomalies primarily focus on static latency analysis. Machine learning (ML) is a potential solution to this problem, as it can support variations in mapping and workloads. However, state-of-the-art ML detectors use costly techniques, usually based on artificial neural networks, adopt synthetic NoC traffic, and use high-level simulations, which do not accurately reflect actual systems. This paper introduces an ML-based flow for detecting traffic anomalies using XGBoost regression, trained on a dataset that captures various traffic characteristics, such as latency and hop count, and different mappings, using a manycore modeled at the RTL level. Our flow presents high accuracy in identifying deviations from expected traffic behavior. This work enhances the security of manycores and opens new opportunities for real-time monitoring and traffic anomaly detection.