Two novel style-transfer palmprint reconstruction attacks

Ziyuan Yang, Lu Leng, Bob Zhang, Ming Li, Jun Chu

Abstract: Palmprint has been widely used for personal authentication in many applications, such that the assessment of recognition system security is important. Online attacks of palmprint recognition are much more difficult than offline attacks due to the fewer permissible login and authentication attempts, the unusability of the matching scores, and less training data. A cross-database attack is another challenging problem, where the images reconstructed from a template can still be effective in attacking the systems with other templates. To achieve online cross-database attacks and ensure that the reconstructed images are high-quality, two novel style-transfer methods are proposed to attack coding-based palmprint recognition systems. The two methods are both based on a convolutional neural network, but their optimization objects are different. In the first method, the optimization object is the input image, where a high-quality image can be reconstructed from the binary template. In the second method, the style-transfer neural network is trained with a template dataset and only one style image to reduce the style loss between the source and target domains. The trained style-transfer network can reconstruct approximately 270 images per second. The two methods have highly impressive attack success rates and satisfactorily meet the requirements of the evaluation system.