Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Assessing the Use of Face Swapping Methods as Face Anonymizers in Videos

Mustafa İzzet Muştu, Hazım Kemal Ekenel

TL;DR

This paper tackles privacy preservation in video data by evaluating face swapping methods that replace a target face with a synthetic source to anonymize identities while striving to maintain data utility. It introduces a unified evaluation pipeline combining six face swapping models (SimSwap, REFace, FaceDancer, E4S, G2Face, FAMS) and two anonymization baselines, using synthetic sources from $aifaces$ and the FaceForensics++ dataset, with metrics spanning $SSIM$, $FVD$, ArcFace embeddings, $L_2$ landmark/pose distances, ID similarity, and identity retrieval. The findings reveal that SimSwap and FaceDancer achieve the best temporal consistency and content quality, whereas REFace and G2Face offer stronger anonymization at the expense of realism and stability; E4S and FAMS show weaker performance in video contexts. The work demonstrates the viability of diffusion- and GAN-based face swapping for privacy-preserving video pipelines and highlights a trade-off between anonymization strength and perceptual fidelity, setting the stage for real-time and controllable anonymization developments in future research.

Abstract

The increasing demand for large-scale visual data, coupled with strict privacy regulations, has driven research into anonymization methods that hide personal identities without seriously degrading data quality. In this paper, we explore the potential of face swapping methods to preserve privacy in video data. Through extensive evaluations focusing on temporal consistency, anonymity strength, and visual fidelity, we find that face swapping techniques can produce consistent facial transitions and effectively hide identities. These results underscore the suitability of face swapping for privacy-preserving video applications and lay the groundwork for future advancements in anonymization focused face-swapping models.

Assessing the Use of Face Swapping Methods as Face Anonymizers in Videos

TL;DR

This paper tackles privacy preservation in video data by evaluating face swapping methods that replace a target face with a synthetic source to anonymize identities while striving to maintain data utility. It introduces a unified evaluation pipeline combining six face swapping models (SimSwap, REFace, FaceDancer, E4S, G2Face, FAMS) and two anonymization baselines, using synthetic sources from and the FaceForensics++ dataset, with metrics spanning , , ArcFace embeddings, landmark/pose distances, ID similarity, and identity retrieval. The findings reveal that SimSwap and FaceDancer achieve the best temporal consistency and content quality, whereas REFace and G2Face offer stronger anonymization at the expense of realism and stability; E4S and FAMS show weaker performance in video contexts. The work demonstrates the viability of diffusion- and GAN-based face swapping for privacy-preserving video pipelines and highlights a trade-off between anonymization strength and perceptual fidelity, setting the stage for real-time and controllable anonymization developments in future research.

Abstract

The increasing demand for large-scale visual data, coupled with strict privacy regulations, has driven research into anonymization methods that hide personal identities without seriously degrading data quality. In this paper, we explore the potential of face swapping methods to preserve privacy in video data. Through extensive evaluations focusing on temporal consistency, anonymity strength, and visual fidelity, we find that face swapping techniques can produce consistent facial transitions and effectively hide identities. These results underscore the suitability of face swapping for privacy-preserving video applications and lay the groundwork for future advancements in anonymization focused face-swapping models.

Paper Structure

This paper contains 19 sections, 2 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Face Anonymization
  4. Face Swapping
  5. Method
  6. Models
  7. SimSwap
  8. REFace
  9. FaceDancer
  10. E4S
  11. G2Face
  12. Face Anonymization Made Simple (FAMS)
  13. Evaluation Pipeline
  14. Evaluation Metrics
  15. Temporal consistency
  16. ...and 4 more sections

Figures (2)

  • Figure 1: Proposed face anonymization pipeline.
  • Figure 2: Qualitative anonymization results. The first row shows consecutive frames from an original video in FaceForensics++ dataset. The following four rows show the results of face swapping methods and the same synthetic source face image. The last two rows are the results of face anonymization methods.