Manipulable, reversible and diversified de-identification via face identity disentanglement

Face de-identification has always been a focal point in privacy-preserving research. Most existing de-identification methods focus only on the anonymization phase, neglecting the importance of deanonymization. Moreover, existing reversible de-identification methods are unsatisfactory in terms of diversity and manipulability. To overcome these limitations, we propose MRDD-FID, short for Manipulable, Reversible and Diversified De-identification via Face Identity Disentanglement. The framework realizes individual modification of identity representations while keeping non-identity representations unchanged through face identity disentanglement. Randomized passwords are used for identity modification, thus ensuring complete randomness in the modification process. By utilizing Generative Adversarial Networks (GANs) for training, we effectively enhance the realism and diversity of de-identification. Furthermore, MRDD-FID can precisely control the degree and direction of de-identification based on user-specified strengths and styles without compromising the image quality. Compared to existing methods, MRDD-FID offers higher flexibility and security. Extensive experiments demonstrate the effectiveness and superiority of our method in terms of anonymity, diversity, reversibility and manipulability.