New Hardware-Software Approach Boosts AI Safety on Embedded Systems

Deploying advanced AI, such as deep neural networks, on embedded systems presents significant challenges, especially for safety-critical applications like medical devices. Current hardware solutions prioritize throughput over computational correctness and security, making them vulnerable to fault injection attacks. Software methods for porting algorithms, like quantization, are either too static and inefficient or dynamic but lack the necessary soundness for critical tasks. This paper introduces a comprehensive solution to these issues. It combines a new real-time, dynamic, and sound adaptive-precision quantization approach with specialized hardware. The software component uses left-to-right arithmetic to process the most significant bits first, dynamically adjusting precision and performing sensitivity analysis to manage decision-boundary risks. Complementing this, a novel systolic array hardware design is proposed to execute this left-to-right arithmetic efficiently. Together, this integrated scheme aims to enable resource-efficient and robust AI at the edge, providing sound, high-precision mathematics that is resilient to bit-flip attacks on critical data. The software implementation is complete, with hardware development underway.