Attack and anomaly prediction in networks-on-chip of multiprocessor system-on-chip-based IoT utilizing machine learning approaches

The proliferation of multiprocessor system-on-chip (MPSoC) architectures within the Internet of Things (IoT) has introduced notable security challenges. These architectures' distributed nature, required for smooth communications between the IP cores, opens them up to potential attacks. Among the most significant issues is increased vulnerability to denial-of-service (DoS) attacks in IoT-based MPSoCs, which can influence their functional capabilities and performance. The situation with this vulnerability is further aggravated by the growing tendency to use third-party IPs in MPSoC designs, relying on global supply chains that can provide necessary performance levels as requirements change. This paper addresses the securing network-on-chip (NoC)-based MPSoCs in an IoT environment, targeting their inherent vulnerability due to third-party IP (3PIP) implementation. For timely prediction, in order to prevent potential security risks, particularly DoS attacks, the authors introduce a runtime monitoring mechanism based on ML. The proposed methodology includes the static training of ML models in a smart-placed strategic manner used for runtime attack detection with nominal performance loss. This analysis examines diverse ML models and features in detail, trying to define the use of machine learning for DoS attack identification. The results show considerable differences between the model performances, with XGBoost showing better accuracy, thus implying its potential as a strong solution. Even though the Nive Bayes model shows some degree of reduced accuracy, it implies that a wise selection of models is critical for strengthening the security position in IoT-based MPSoCs.