VerA: Versatile Anonymization Applicable to Clinical Facial Photographs

TL;DR

VerA is presented, the first Versatile Anonymization framework that solves two challenges in clinical applications and reaches the state of the art on established anonymization tasks, in terms of photorealism and de-identification.

Abstract

The demand for privacy in facial image dissemination is gaining ground internationally, echoed by the proliferation of regulations such as GDPR, DPDPA, CCPA, PIPL, and APPI. While recent advances in anonymization surpass pixelation or blur methods, additional constraints to the task pose challenges. Largely unaddressed by current anonymization methods are clinical images and pairs of before-and-after clinical images illustrating facial medical interventions, e.g., facial surgeries or dental procedures. We present VerA, the first Versatile Anonymization framework that solves two challenges in clinical applications: A) it preserves selected semantic areas (e.g., mouth region) to show medical intervention results, that is, anonymization is only applied to the areas outside the preserved area; and B) it produces anonymized images with consistent personal identity across multiple photographs, which is crucial for anonymizing photographs of the same person taken before and after a clinical intervention. We validate our results on both single and paired anonymization of clinical images through extensive quantitative and qualitative evaluation. We also demonstrate that VerA reaches the state of the art on established anonymization tasks, in terms of photorealism and de-identification.

Figures (21)

• Figure 1: Clinical paired image anonymization. VerA anonymizes two photographs of a person before (left) and after (right) a clinical intervention in the mouth. Top row: the input image pair with a common identity (the same person, but different photographs) and a semantic region to preserve, which is the medically treated area. Second and third row: two example outputs with each a before- (left) and an after-intervention image (right). In the example outputs, the faces except the preserved areas are completely anonymized while the persons' identities in the output images before and after the interventions are preserved.
• Figure 2: VerA exploits a novel controllable image generator with specialized inversion designed for anonymization tasks, unlike previous methods built on pretrained generators. Compared to SemanticStyleGAN shi2022semanticstylegan that can only control semantics, our illustrated generator has a dual latent space (z$_g$) with individual mappings for high-level attributes. We adapt the conditioning layers of each generator to the dual space, and design a generative contrastive learning approach to learn the high-level controls.
• Figure 3: Illustration of accumulated semantic and high-level edits (left to right modifications: hair, pose, eyes, and age) with our trained image generator.
• Figure 4: The effect of segmentation loss on inversion. (a) Input and predicted segmentation, (b) inversion results without and (c) with segmentation loss. Overlaid edges show the reference segmentation from the input (a), illustrating improved inversion.
• Figure 5: Our clinical single-image anonymization results preserving respectively the mouth, nose, or eyes, compared with state-of-the-art anonymization methods.