Detecting Adversarial Audio via Activation Quantization Error

The robustness and vulnerability of Deep Neural Networks (DNN) are quickly becoming a critical area of interest since these models are in widespread use across real-world applications (i.e., image and audio analysis, recommendation system, natural language analysis, etc.). A DNN's vulnerability is exploited by an adversary to generate data to attack the model; however, the majority of adversarial data generators have focused on image domains with far fewer work on audio domains. More recently, audio analysis models were shown to be vulnerable to adversarial audio examples (e.g., speech command classification, automatic speech recognition, etc.). Thus, one urgent open problem is to detect adversarial audio reliably. In this contribution, we incorporate a separate and yet related DNN technique to detect adversarial audio, namely model quantization. Then we propose an algorithm to detect adversarial audio by using a DNN's quantization error. Specifically, we demonstrate that adversarial audio typically exhibits a larger activation quantization error than benign audio. The quantization error is measured using character error rates. We use the difference in errors to discriminate adversarial audio. Experiments with three the-state-of-the-art audio attack algorithms against the DeepSpeech model show our detection algorithm achieved high accuracy on the Mozilla dataset.