Context-Aware Full Body Anonymization using Text-to-Image Diffusion Models

TL;DR

This paper proposes a workflow for full body person anonymization utilizing Stable Diffusion as a generative backend and shows that the method outperforms state-of-the art anonymization pipelines with respect to image quality, resolution, Inception Score (IS) and Frechet Inception Distance (FID).

Abstract

Anonymization plays a key role in protecting sensible information of individuals in real world datasets. Self-driving cars for example need high resolution facial features to track people and their viewing direction to predict future behaviour and react accordingly. In order to protect people's privacy whilst keeping important features in the dataset, it is important to replace the full body of a person with a highly detailed anonymized one. In contrast to doing face anonymization, full body replacement decreases the ability of recognizing people by their hairstyle or clothes. In this paper, we propose a workflow for full body person anonymization utilizing Stable Diffusion as a generative backend. Text-to-image diffusion models, like Stable Diffusion, OpenAI's DALL-E or Midjourney, have become very popular in recent time, being able to create photorealistic images from a single text prompt. We show that our method outperforms state-of-the art anonymization pipelines with respect to image quality, resolution, Inception Score (IS) and Frechet Inception Distance (FID). Additionally, our method is invariant with respect to the image generator and thus able to be used with the latest models available.

Figures (11)

• Figure 1: Images from different scenes anonymized with the method proposed in this paper. lin2014microsoft
• Figure 2: The Markov chain used for DMs. The initial state $x_0$ is corrupted by iteratively adding noise $q(x_{n+1} | x_n)$ until arriving at the fully noised image $x_N$.
• Figure 3: A high level overview of our anonymization pipeline.
• Figure 4: How the parameter $\beta_d$ influences the anonymization strength of a person instance.
• Figure 5: Comparison of DeepPrivacy2 and our method on a high resolution stock photo witbooi2023.