Adversarial Attacks on Vehicle Re-Identification

Integral to intelligent transportation, vehicle re-identification (re-ID) involves a match of the same vehicle across different cameras and therefore has a vital role to play in monitoring and security. Recently, the application of deep learning to computer vision has drawn great attention, inspired by which, deep learning-based vehicle re-ID models have gained popularity and delivered great outcomes. However, it has been proved that a deep neural network could face the problem of adversarial examples, which render a model significantly vulnerable by adding perturbations to an image. Although there are many researches on adversarial attacks on classification tasks, few of them focus on the safety of vehicle re-ID, whose algorithm, as a deep learning algorithm, is most likely to grapple with security flaws. Such attacks pose grave risks as they can barely be detected by humans. A metric attack on the vehicle means modifying the distance between features of attacked images and those of reference images. In this paper, we leveraged the baselineNet, a vehicle re-ID network, to make a comparison among multiple metric learning-based attacks and proposed the Average Furthest-Negative Attack (AFNA) algorithm, which can produce better results at a lower cost of time in re-identifying vehicles. Effects of attacks were evaluated on the VeRi-776 dataset.