Adversarial Examples in Self-Driving: A Review of Available Datasets and Attacks

Abstract

Autonomous vehicles rely on computer vision models for perception, which have been shown to be vulnerable to adversarial attacks. These attacks pose various risks from reducing user confidence in the technology to directly influencing the technology to make a particular action [1]. Research in adversarial machine learning (ML) has led to increased awareness of ML’s vulnerabilities to such attacks and strategies to mitigate their effects and enhance ML generalizability [1]. Although there exists a dataset for assessing a computer vision model’s efficacy against adversarial attacks in the real-world [2], there is no such dataset for benchmarking computer vision algorithms for autonomous vehicles. This effort is a preliminary study on publicly available datasets for self-driving applications through the lens of adversarial ML. We include a discussion of attacks that have been performed in this space. From black-box to white-box attacks, the more knowledge is available to the adversary, the more they can leverage to employ the attack [1]. Adversaries are fully capable of utilizing the publicly available datasets to test their own black-box attacks [3], [4]. Adversarial examples can be useful in both enhancing algorithm resilience to these attacks through inclusion in training [5] as well as for use in benchmarking current approaches against the attacks [2]. Building resilient autonomy in transportation will require mitigation of any vulnerabilities, especially those from potential adversaries. This survey of datasets and attacks on self-driving vehicles is a first step in developing a dataset of adversarial attacks in this domain. The dataset can assist current and future computer vision research in self-driving to benchmark algorithm effectiveness against such attacks in the real-world.