Undermining License Plate Recognition: A Data Poisoning Attack

With the development of computer vision, License Plate Recognition (LPR) systems are widely employed in various applications, including traffic management and intelligent transportation. However, recent studies have shown that these systems are vulnerable to adversarial attacks, which significantly undermine their accuracy and reliability. Existing works typically fall into evasion attacks and model poisoning attacks, which often lack feasibility and stealthiness. In contrast, data poisoning attacks have the advantage of being more practical and effective in real-world scenarios. In this paper, we propose a novel data poisoning attack against LPR systems, which leverages gradient matching to generate poison samples that can mislead the recognition module of the systems. Notably, we focus on a challenging scenario, where we control only a small fraction of training data and attack a model trained from scratch without being perceptible. Moreover, we conduct experiments to validate our attack on the benchmark dataset CCPD2019 and demonstrate that our attack achieved outstanding performance by manipulating only 1% of the training data with perturbations less than 8 pixels in ℓ
∞
-norm. Our work highlights the potential vulnerabilities of LPR systems and encourages research into more trustworthy systems.