DeepPrivacy: A Generative Adversarial Network for Face Anonymization (1909.04538v1)

Abstract: We propose a novel architecture which is able to automatically anonymize faces in images while retaining the original data distribution. We ensure total anonymization of all faces in an image by generating images exclusively on privacy-safe information. Our model is based on a conditional generative adversarial network, generating images considering the original pose and image background. The conditional information enables us to generate highly realistic faces with a seamless transition between the generated face and the existing background. Furthermore, we introduce a diverse dataset of human faces, including unconventional poses, occluded faces, and a vast variability in backgrounds. Finally, we present experimental results reflecting the capability of our model to anonymize images while preserving the data distribution, making the data suitable for further training of deep learning models. As far as we know, no other solution has been proposed that guarantees the anonymization of faces while generating realistic images.

• The paper introduces DeepPrivacy, a Conditional Generative Adversarial Network (GAN) that anonymizes faces while preserving the original data distribution.
• DeepPrivacy achieves a negligible degradation in face detection performance, retaining 99.3% average precision on the WIDER-Face dataset, and presents the 1.47 million face Flickr Diverse Faces (FDF) dataset.
• Detailed ablation studies confirm that specific architectural choices are crucial for retaining data distribution, demonstrating the method's superiority over simpler anonymization techniques.

DeepPrivacy: A Generative Adversarial Network for Face Anonymization

This paper introduces DeepPrivacy, a Conditional Generative Adversarial Network (GAN) designed to anonymize faces within images while preserving their original data distribution. The underlying motivation for this work emerges from the growing need for privacy-preserving data handling methods, especially with the enforcement of legal frameworks like the General Data Protection Regulation (GDPR). DeepPrivacy presents a model that aptly balances the task of face de-identification without compromising the natural statistics of the input data, a challenge that existing traditional methods have not adequately addressed.

The architectural design of DeepPrivacy is noteworthy. It leverages a conditional GAN architecture, specifically considering the image background and original pose to achieve a seamless transition between the generated face and the image's remaining components. This approach ensures the generation of realistic visual outputs without exposure to privacy-sensitive data. The model is built upon a U-net architecture, enhanced by the progressive growing training methodology, initiating at a resolution of 8x8 and elevating to 128x128. This progressively larger scale allows the GAN to maintain high training efficiency and image quality.

A significant contribution of this research is the introduction of the Flickr Diverse Faces (FDF) dataset, containing 1.47 million annotated faces. The FDF dataset showcases a comprehensive variety of poses, occlusions, and backgrounds, providing a diverse training ground for DeepPrivacy, and improving its robustness in real-world applications.

One distinguishable achievement of this method is its impact on face detection tasks. In comparison studies using the WIDER-Face dataset, the model demonstrated a negligible degradation in face detection performance, ensuring 99.3% of the average precision of the original dataset. This enhances the argument for its applicability in scenarios where user privacy needs de-identification without losing analytic potential.

Furthermore, detailed ablation experiments affirm the architectural choices made by the authors. Specifically, the inclusion of sparse pose data and the selection of a wider discriminator with no normalization layer adjustments show superior performance in retaining data distribution. The GAN structure effectively surpasses more straightforward anonymization methods such as pixelation or blurring, revealing the inadequacies of those techniques in maintaining both privacy and analytic utility.

Despite its results, DeepPrivacy does present limitations, particularly in handling highly occluded faces or complicated backgrounds, where the model occasionally generates unrealistic images. These areas signify potential directions for future research, suggesting improvements such as denser pose estimations or enhancements in training data quality to better handle visual complexities.

Conclusively, DeepPrivacy stands as a potent model for anonymization in visual datasets, merging the needs for privacy with the demands of high-quality image generation. Its design is scalable and adaptable, setting a precedent for integrating more sophisticated methods capable of managing complex, privacy-preserving tasks without compromising the analytic integrity of the data. This foundational work paves the way for future developments in AI, focusing on achieving ethical data handling practices in large-scale deployments.