Hardware Trojan Key-Corruption Detection with Automated Neural Architecture Search

This work presents a model hardware trojan which intermittently is capable of corrupting an encryption operation occurring on a device. It asks whether this trojan can be detected via power-based, side-channel attacks only instrumenting the encryption itself, not the control flow of the trojan itself. By applying Automated Machine Learning techniques to search neural architecture, a classification of corrupted encryption operations is able to completely identify whether the operation corresponded with a corrupted operation or not. Through a number of experiments, we demonstrate this fact holds regardless of variable or constant plaintext, rotating encryption keys, or even with different corrupted keys.