[Paper Review] CFA-Net: Controllable Face Anonymization Network with Identity Representation Manipulation

Tianxiang Ma, Dongze Li+2 more|arXiv (Cornell University)|May 24, 2021

Face recognition and analysis43 references11 citations

TL;DR

CFA-Net proposes a controllable face anonymization framework that disentangles identity from identity-uncorrelated attributes using a disentangled representation space, enabling flexible and precise anonymization by manipulating identity vectors. It achieves state-of-the-art performance in anonymization effectiveness, image quality, and controllability, with FID of 2.193 and 98.8% anonymization success rate on CelebA-HQ.

ABSTRACT

De-identification of face data has drawn increasing attention in recent years. It is important to protect people's identities meanwhile keeping the utility of the data in many computer vision tasks. We propose a Controllable Face Anonymization Network (CFA-Net), a novel approach that can anonymize the identity of given faces in images and videos, based on a generator that can disentangle face identity from other image contents. We reach the goal of controllable face anonymization through manipulating identity vectors in the generator's identity representation space. Various anonymized faces deriving from an original face can be generated through our method and maintain high similarity to the original image contents. Quantitative and qualitative results demonstrate our method's superiority over literature models on visual quality and anonymization validity.

Motivation & Objective

• To address the lack of controllability in existing face anonymization methods, which often produce fixed or low-quality outputs.
• To preserve identity-uncorrelated attributes such as pose, expression, and haircut while effectively anonymizing identity features.
• To ensure anonymized faces do not leak identity information of other individuals in the database.
• To enable flexible, user-controlled anonymization with adjustable levels of identity modification.
• To decouple the training and anonymization phases, eliminating dependency on reference identities or additional supervision.

Proposed method

• The method employs a generator with a disentangled identity representation space, where identity and content are separated through adversarial training.
• Identity disentanglement is achieved via a dual-branch network with identity-specific and content-specific encoders, trained using identity and content reconstruction losses.
• Anonymization is performed by sampling and modifying the original identity vector in the disentangled space while keeping content vectors fixed.
• The model uses a content consistency loss to preserve texture, color, and structural details of the original face.
• An identity loss is applied to ensure the model learns a meaningful identity representation that can be manipulated for anonymization.
• The framework is trained end-to-end on a large-scale face dataset, enabling zero-shot anonymization of unseen faces without reference images.

Experimental results

Research questions

• RQ1 Can face anonymization be made controllable through direct manipulation of identity vectors in a disentangled representation space?
• RQ2 To what extent can identity-uncorrelated attributes like pose and expression be preserved during anonymization?
• RQ3 How does the proposed method compare to prior works in anonymization success rate and image quality metrics?
• RQ4 What is the impact of key loss functions (content and identity loss) on the quality and effectiveness of anonymization?
• RQ5 Can the method be effectively applied to video sequences while maintaining temporal consistency?

Key findings

• CFA-Net achieves a Fréchet Inception Distance (FID) of 2.193 on CelebA-HQ, significantly outperforming prior methods (e.g., 2.663 for Maximov et al.).
• The model achieves a low LPIPS score of 0.177, indicating high perceptual similarity to the original images, compared to 0.221 in prior work.
• CFA-Net achieves 98.8% anonymization success rate, surpassing the 93.7% of the ablation without content loss and the 0.4% without identity loss.
• The method maintains high SSIM (0.865) and PSNR (21.26) within detected face regions, indicating strong structural and pixel-level fidelity.
• Face detection rates remain high across multiple detectors (MTCNN: 92.0%, Dlib: 95.9%, SSH: 100%), confirming utility for downstream tasks.
• Ablation studies confirm that both content and identity losses are essential: removing either leads to severe degradation in image quality or anonymization failure.