Adversarial Perturbations for License Plate Information Privacy

In an era where data privacy is increasingly important, the public release of large datasets can greatly benefit the research community while also raising concerns about the misuse of private information. To balance these interests, one solution is to empower the authorities to grant only authorized models access to usable but sensitive information within a dataset. For instance, in scenarios involving the public release of vehicle images, it is crucial to de-identify sensitive information such as license plates while maintaining the data’s utility for law enforcement agencies. This study develops a novel method for protecting the privacy of data when it is publicly released while ensuring that essential information remains accessible for authorized operations. Our approach involves de-identifying image sections containing private information and then applying adversarial modifications to preserve useful information exclusively for certain authentication models. Experiments demonstrate that, after applying adversarial attacks, authentication models can re-identify information on de-identified data with up to accuracy of 94.8% for detection and 80% for OCR, whereas non-authentication models are mostly unable to recognize the information, with 0.1% and 0.25% for score of detection and OCR respectively. This research highlights the potential of adversarial attacks in enhancing privacy protection, particularly in the context of license plate recognition.