Blurring Boundaries: A New Way to Secure Approximate Computing Systems
Pruthvy Yellu, Landon Buell, Dongpeng Xu, Qiaoyan Yu · 2020
Approximate computing (AC) techniques have been widely used to improve the performance of computing systems by trading off accuracy. However, recent literature projects that the utilization of approximation could bring in new security threats to computing systems. This work presents two practical attacks on the AC systems for multilayer perceptron (MLP) and Sobel algorithm based image edge detection. The case studies in this work indicate that the approximation mechanism in AC systems can be exploited to conduct stealthy attacks, which suddenly cause significant degradation in accuracy and lead to unpredictable primary outputs. To address the emerging threats on AC systems, this work proposes to blur the boundary between approximate and precise computing submodules in AC systems. This new defense method obscures that boundary with three obfuscation schemes such that adversary could not easily identify the right target to precisely perform hardware tampering attacks. Simulation results show that the proposed method can effectively reduce the attack success rate.