Excellence in Research: A Hierarchical Machine Learning Approach for Securing of NoC-Based MPSoCs Against Thermal Attacks

点击翻译按钮获取中文摘要

The design of Multi-Processor System-on-Chips (MPSoCs) often involves the integration of pre-designed Intellectual Property (IP) components to minimize costs and accelerate time to market. This approach leaves room for potential manipulation of the manufacturing process by adversaries who insert malicious circuitries known as Hardware Trojans (HTs) into the final product. Depending on the intentions of the adversary, an HT can perform various malicious tasks, including compromising reliability, causing operational failures, leaking information, and initiating denial of services. This project aims to address security concerns related to HT-infected thermal sensors embedded in MPSoCs. Given that thermal information is notably used in dynamic power and thermal management, it is crucial to monitor the behavior of thermal sensors within an MPSoC to detect and isolate compromised ones. This project aims to achieve this goal by employing a hierarchical machine learning (ML) approach. This project impacts a broad range of computing systems that utilize any of the commercially available MPSoCs on the market.In order to monitor the functionality of thermal sensors in an MPSoC, the thermal information obtained from the cores on the chip undergoes processing through a hierarchy of small to complex machine learning (ML) classifiers. At the lowest level, countermeasures implemented at the Network-on-Chip (NoC) routers within the target MPSoC try to identify compromised thermal sensors. The thermal data collected by each router is then transmitted to a chip-wide ML classifier, which functions as a dedicated ML accelerator, capable of capturing cases that are not easily detected by the router-level countermeasures. Subsequently, the thermal data is often transmitted to a cloud server for further ML processing, serving as a feedback mechanism to update the weights of the on-chip ML classifier. As the accuracy of the on-chip classifier improves through learning feedback from the cloud-based classifier, the proposed approach has the potential to address attacks with diverse probabilistic characteristics and profiles.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.