Optimized AI for Secure and Efficient Embedded System
Zesheng Li, Yuan Pan · 2024
With the increased penetration of embedded systems in different areas such as IoT devices, industrial control systems and autonomous vehicles, the issue of security of such systems is of paramount importance. These systems are generally constrained by limited computational resources and energy availability in order to run on a reasonable battery size and duration and therefore the use of traditional security solutions is not feasible on such platforms. In this paper, we show the development and optimisation of AI algorithms specifically catered for detection of real-time threats and defence of embedded systems with limited computational and energy resources. We explore light-weight machine learning models optimised for latency, speed and power-awareness. Additionally, we employ innovative methods such as machine learning model quantisation and pruning to allow these algorithms to run on low-power processors. We validate the methodologies using case studies as well as an assessment of the performance of the algorithms with regards to classification accuracy, energy consumption and latency. Our AI-based solutions demonstrate superior detection accuracy, reduced energy consumption and minimised response times compared with state-of-the-art solutions, showing that these provide a balanced approach to embedded system security, which can be deployed where security as well as resource efficiency are paramount.