Can We Securely Use Approximate Computing?

Approximate computing (AC) techniques bring in an alternative way to improve energy efficiency for computing systems, at the cost of acceptable reduction on accuracy. Unfortunately, recent literature indicates that approximate computing systems could be vulnerable to new security threats. Approximation mechanisms maybe leveraged to introduce more errors than the level that the AC systems can tolerate. This work analyzes the existing metrics for accuracy from attack detection point of view. A differential metric is proposed to enlarge the Trojan induced difference on accuracy, delay and power, thus easing Trojan detection. Furthermore, this work proposes a unique attack detection method for AC systems to reduce false negative rate on Trojan detection. Our case study shows that the proposed method can reduce the false negative rate by 93%.