Joint reconstruction and deidentification for mobile identity anonymization

The growing use of deep learning methods in various applications has raised concerns about privacy, as these methods heavily rely on large-scale datasets. To address this issue, anonymization methods have been proposed that usually use face detection models to remove the original face and regenerate a new one. Although these methods have achieved promising performance, alternative face detection models significantly increase the training cost and inference time. To solve these problems, we propose an end-to-end mobile anonymization method with a joint reconstruction and deidentification (JRD) framework. We introduce deidentification loss into a generative adversarial network to anonymize the original image and add reconstruction loss to reconstruct the background of the original image. By balancing these two losses, we can anonymize the identity without changing the background and no longer rely on additional face detection models. In addition, to ensure the essential difference between the generated and original images, we generated a random code for each original image. Further, we verified the effectiveness and superiority of JRD on the CelebA dataset. The experimental results show that JRD not only outperforms existing one-to-one methods but is also superior to many-to-one methods.