Enhancing thermal attack resilience in Multi-Processor System-on-Chip through synthetic data generation using LLMs

The security of Multi-Processor System-on-Chip (MPSoC) architectures has become a critical concern due to their widespread integration into modern computing systems. However, the lack of quality data capturing the behavior of these systems under attack scenarios hinders the development of effective countermeasures. In this study, we propose a framework that facilitates the collection and enrichment of thermal datasets using machine learning for MPSoC security analysis against two distinct thermal attack patterns. To address the limitations associated with imbalanced data, the proposed framework leverages Large Language Models (LLMs) to generate synthetic data, incorporating a feedback-driven optimization mechanism to enhance data quality and alignment with real-world distributions, we developed a prompt-based LLM pipeline for schema-conformant tabular synthesis and evaluate its downstream utility by training standard off-the-shelf ML models on the generated tables. The framework employs a novel prompting method to ensure compatibility across various LLMs, thereby optimizing the synthetic data generation process. Experimental evaluations provide empirical evidence substantiating the framework’s efficiency in generating and enriching synthetic data for training machine learning models. The experimental evaluation reveals that the Light Gradient Boosting Machine algorithm achieved the best performance among all tested models, attaining an accuracy of 93% and an F1-score of 92.80% when trained only on synthetic data and evaluated on real data. The findings highlight its potential as a powerful tool for strengthening MPSoC security by improving the robustness and adaptability of machine learning-based security mechanisms.