My Face My Choice: Privacy Enhancing Deepfakes for Social Media Anonymization

View on arXiv ← Back to list

Authors: Umur A. Ciftci, Gokturk Yuksek, Ilke Demir

Published: 2022-11-02 17:58:20+00:00

AI Summary

This research introduces a privacy-enhancing system, "My Face My Choice" (MFMC), that replaces unapproved faces in social media photos with quantitatively dissimilar deepfakes. The system utilizes face embeddings and diverse deepfake generators to ensure anonymity while maintaining image quality, significantly reducing the accuracy of face recognition systems.

Abstract

Recently, productization of face recognition and identification algorithms have become the most controversial topic about ethical AI. As new policies around digital identities are formed, we introduce three face access models in a hypothetical social network, where the user has the power to only appear in photos they approve. Our approach eclipses current tagging systems and replaces unapproved faces with quantitatively dissimilar deepfakes. In addition, we propose new metrics specific for this task, where the deepfake is generated at random with a guaranteed dissimilarity. We explain access models based on strictness of the data flow, and discuss impact of each model on privacy, usability, and performance. We evaluate our system on Facial Descriptor Dataset as the real dataset, and two synthetic datasets with random and equal class distributions. Running seven SOTA face recognizers on our results, MFMC reduces the average accuracy by 61%. Lastly, we extensively analyze similarity metrics, deepfake generators, and datasets in structural, visual, and generative spaces; supporting the design choices and verifying the quality.

Key findings

MFMC reduces the average accuracy of seven state-of-the-art face recognizers by 61%. SimSwap is identified as the best deepfake generator. The choice of similarity metric for selecting deepfakes impacts the results.

Approach

MFMC replaces unapproved faces with deepfakes generated from a pool of synthetic faces. Face embeddings are used to select dissimilar deepfakes, ensuring anonymity while preserving image context. Three access models are proposed to manage privacy levels.

Datasets

Facial Descriptor Dataset (real dataset), two synthetic datasets (StyleGAN with random distribution, Generated.Photos with equal distribution across skin tones and genders)

Model(s)

InsightFace (for face detection and embeddings), SimSwap (for deepfake generation), ArcFace (for embedding extraction), and several other GANs (FTGAN, FSGAN, FaceSwap) for comparison.

Author countries

USA, USA

← Previous