Disentangle Before Anonymize: A Two-stage Framework for Attribute-preserved and Occlusion-robust De-identification

TL;DR

Disentangle Before Anonymize (DBAF) tackles privacy-preserving face de-identification by separating identity disentanglement from anonymization in two stages. It introduces CID for contrastive identity disentanglement, KRIA for key-based reversible anonymization, and MAAR for maintaining attribute fidelity under occlusion, built on an E4E encoder and StyleGAN2. The two-stage design yields more faithful attribute preservation, higher-quality anonymized outputs, and stronger occlusion robustness than prior reversible methods. Extensive experiments on FFHQ and CelebA-HQ show superior anonymization and recovery performance, plus diverse identity variation and preserved background details. This approach offers reversible, high-fidelity de-identification suitable for privacy protection in real-world photo-sharing.

Abstract

In an era where personal photos are easily leaked and collected, face de-identification is a crucial method for protecting identity privacy. However, current face de-identification techniques face challenges in preserving attribute details and often produce anonymized results with reduced authenticity. These shortcomings are particularly evident when handling occlusions,frequently resulting in noticeable editing artifacts. Our primary finding in this work is that simultaneous training of identity disentanglement and anonymization hinders their respective effectiveness.Therefore, we propose "Disentangle Before Anonymize",a novel two-stage Framework(DBAF)designed for attributepreserved and occlusion-robust de-identification. This framework includes a Contrastive Identity Disentanglement (CID) module and a Key-authorized Reversible Identity Anonymization (KRIA) module, achieving faithful attribute preservation and high-quality identity anonymization edits. Additionally, we introduce a Multiscale Attentional Attribute Retention (MAAR) module to address the issue of reduced anonymization quality under occlusions.Extensive experiments demonstrate that our method outperforms state-of-the-art de-identification approaches, delivering superior quality, enhanced detail fidelity, improved attribute preservation performance, and greater robustness to occlusions.

Figures (13)

• Figure 1: De-identification and recovery results generated by RiDDLE li2023riddle, FALCO barattin2023attribute, and our method. Besides being able to synthesize high-fidelity anonymized images, DBAF has obvious advantages in attribute preservation and occlusion robustness.
• Figure 2: The anonymization and recovery process of DBAF. In the anonymization phase, given the input face image $X_{ori}$ and a randomly generated key $P$, DBAF transforms the input face into an anonymized version. During the recovery phase, if the entered key matches the one generated in the anonymization process, DBAF generates an image with the original identity. Conversely, if the entered key does not match, DBAF produces an image with a different identity.
• Figure 3: Overview of the stages and modular structure of DBFA. On the basis of a pre-trained E4E encoder tov2021designing and a StyleGAN2 decoder karras2020analyzing, DBAF introduces three modules: CID, KRIA, and MAAR. CID and MAAR are involved in both stages of training, whereas KRIA is trained exclusively in the second stage. The two stages utilize different inputs and loss constraints to guide the model's learning effectively.
• Figure 4: The architecture of DBAF. DBAF comprises three key components. The CID module separates the latent code of the input image into an attribute latent code and an identity latent code. The KRIA module edits the identity latent code using key-based controls to achieve identity modifications. The MAAR module preserves and enhances the attribute details, ensuring high fidelity and accurate attribute representation.
• Figure 5: Qualitative de-identification results on CelebA-HQ karras2017progressive. Our method produces high-quality images while preserving key attribute features, including background details and fine textures such as wrinkles, more effectively than existing approaches.