Investigating the Impact of Adversarial Attacks on Deep Learning-Based Wearable Robot Controllers: Security, Reliability, and Safety Concerns

Advances in deep learning have motivated its inclusion in wearable robot control to accelerate their translation efforts to real-world environments like industrial, military, healthcare, and community. Yet, vulnerabilities of deep learning-based wearable robot controllers towards crafted attacks that disrupt robot motion to compromise users' and organizations' security and safety remain unexplored. This paper establishes such vulnerabilities in deep learning-based wearable robot controllers by conducting adversarial attacks where indistinguishable adversarial input signals are developed by adding minor perturbations in original kinematic input signals. We propose two novel and biomechanically informed adversarial attack methods that force deep learning models to estimate the gait phase chosen by adversaries, consequently disrupting robot motion. These attacks instigated a significant drop in prediction accuracy of multiple popular deep learning models used to estimate individuals' gait phase. Mean absolute error (MAE) increased from 2.87 up to 50.43 for Convolutional Neural Networks and from 1.88 up to 50.04 for Long-Short Term Memory Networks. Consequently, this significantly altered the Hip, Knee, and Ankle flexion profiles from original (up to 49.55 Nm). Yet the original and adversarial samples had a high similarity of 92.5% and above. Thus, our attacks seamlessly altered the behavior of deep learning-based wearable robot controllers as per adversaries' intention, consequently compromising security, reliability, and safety. Our study is the first to underscore the emerging risks of adversarial attacks on deep learning-based wearable robot controllers, serving as a starting point for future studies on safeguarding them against real-world attacks where ramifications of disrupted robot motion are detrimental.