Blurring Boundaries: A New Way to Secure Approximate Computing Systems

Security Threat Analyses and Attack Models for Approximate Computing Systems

“…Those case studies mentioned above motivate us to perform a systematic study on the hardware attack methods that could harm AC systems. Our recent work [78] performs a preliminary analysis on security threats, which have emerged due to the easier identification of boundaries between precise and approximate modules. In spite of the recent progress that has been made in References [47,52,78,79], a comprehensive study on security challenges brought by various approximation techniques is not available yet.…”

“…Our recent work [78] performs a preliminary analysis on security threats, which have emerged due to the easier identification of boundaries between precise and approximate modules. In spite of the recent progress that has been made in References [47,52,78,79], a comprehensive study on security challenges brought by various approximation techniques is not available yet. The investigation on the security threats on AC systems is still in its youth due to the following reasons: (1) as different abstraction layers have diverse approximation mechanisms, it is difficult to formulate the security problem in a systematic manner; (2) there does not exist a unified framework to standardize the procedure for discovering the security vulnerabilities of AC systems;…”

“…Our recent work [78] presents a new type of attacks in AC systems-Error CompensAtion (ECA) attacks, which take advantage of the relaxed accuracy to partially replace the precise functionality with malicious behaviors. Although the behavior defined in the attack will result in a sudden drop on accuracy/precision, the average accuracy still remains in the acceptable range due to approximation.…”

“…The works [37,74] considers the approximate technique that truncate certain LSBs in a data for power saving and proposes to embed security bits in the truncated data to detect errors. Reference [78] protects AC systems by blurring the boundary between AC and non-AC zones such that the obfuscated boundary can mislead attackers and thus improve the success rate of attack detection. As shown in Figures 14(a) and 14(b), the boundary blurring techniques obfuscate the appearance of AC modules from the AC system users' point of view.…”

“…To address the ECA attack, we applied the WTA scheme [78] to blur the boundary between approximate and precise operations such that the utilization of the approximation module is not explicit and uncertain from the attacker's point of view. In this case study, the first hidden layer of the CNN model used approximate additions in the matrix computation to accelerate the processing speed at the neural nodes.…”