Security Vulnerability (Backdoor Trojan) During Machine Learning Accelerator Design Phases

This article provides a broad perspective for readers on how security vulnerabilities can be exploited by hackers during the design of machine learning (ML) accelerators such as convolutional neural networks. The article also highlights different ML accelerators from academia and industry, along with various design stages where Trojan attacks can be launched. It also provides a broad spectrum review of existing literatures on how Trojan attacks can be injected, triggered, and exploited to cause various payloads, such as degraded performance, denial of service, data damage, and power/battery exhaustion, on ML accelerators. The article also discusses possible detection/mitigation techniques from the literature for some of the attacks and recommends a possible solution.