A Multiversion Programming Inspired Approach to Detecting Audio Adversarial Examples

Abstract

Adversarial examples (AEs) are crafted by adding human-imperceptible perturbations to inputs such that a machine-learning based classifier incorrectly labels them. They have become a severe threat to the trustworthiness of machine learning. While AEs in the image domain have been well studied, audio AEs are less investigated. Recently, multiple techniques are proposed to generate audio AEs, which makes countermeasures against them urgent. Our experiments show that, given an audio AE, the transcription results by Automatic Speech Recognition (ASR) systems differ significantly (that is, poor transferability), as different ASR systems use different architectures, parameters, and training datasets. Based on this fact and inspired by Multiversion Programming, we propose a novel audio AE detection approach MVP-EARS, which utilizes the diverse off-the-shelf ASRs to determine whether an audio is an AE. We build the largest audio AE dataset to our knowledge, and the evaluation shows that the detection accuracy reaches 99.88%.While transferable audio AEs are difficult to generate at this moment, they may become a reality in future. We further adapt the idea above to proactively train the detection system for coping with transferable audio AEs. Thus, the proactive detection system is one giant step ahead of attackers working on transferable AEs.

PFirewall: Semantics-Aware Customizable Data Flow Control for Smart Home Privacy Protection

“…HAWatcher [75] proposes a novel semantics-aware approach that utilizes semantic information, such as automation rules and device relations, to mine correlations from event logs for precise and explainable anomaly and attack detection. Security issues related to voice assistants draw great attention [76], [77], [78]. For example, MVP-Ears [78] presents a highly accurate (accuracy > 99.8%) audio adversarial example (AE) detection method inspired by multiversion programming, and it can proactively handle transferable audio AEs.…”

“…Security issues related to voice assistants draw great attention [76], [77], [78]. For example, MVP-Ears [78] presents a highly accurate (accuracy > 99.8%) audio adversarial example (AE) detection method inspired by multiversion programming, and it can proactively handle transferable audio AEs. This work is concerned with the rich privacy-sensitive data generated by IoT devices.…”

Adversarial Attacks for Image Segmentation on Multiple Lightweight Models

“…Also, recent studies show that adversarial examples can be applied to the real worlds, such as object recognition system [41], controllable voice system [24] and traffic sign recognition system [25]. Zeng et al proposed a novel audio detection approach to determine whether audio is an adversarial example [26]. Xiao et al designed a malware detection scheme with Q-learning for a mobile device to derive the optimal offloading rate.…”

Privacy Leakage in Smart Homes and Its Mitigation: IFTTT as a Case Study

“…Although manufacturers have deployed various measurements of their platforms, some papers still found flaws leaking users' privacy on various platforms [18]- [21]. There have been many papers showing how to compromise a user's privacy via the flaws of cloud [22]- [25], protocols, voice interface [26], or even traffic analysis. These efforts cover the privacy breaches that can be caused to users by various platforms in a smart home environment.…”