Introduction
The rapid advancement of deepfake technology poses a significant threat to individual privacy. Current defensive strategies primarily fall into two categories: detection and destruction. Detection is inherently reactive, trailing behind the evolution of synthesis models. Destruction methods, which typically involve injecting adversarial noise into images to disrupt deepfake algorithms, often degrade image quality with visible artifacts and frequently fail to generalize across different, unseen deepfake models.
This paper introduces Scapegoat Generation, a proactive approach that decouples a user's digital persona from their physical identity. Instead of attempting to break the synthesis process, the method transforms the source image into a "scapegoat"—a semantically edited version of the user's face that remains recognizable as an avatar but is mathematically distinct from the original identity.
Methodology: Adversarial Facial Style Modification
The core of the proposed system is a pipeline for "Adversarial Facial Style Modification" using GAN inversion. The process involves three main stages:
- Latent Embedding: Using a pre-trained StyleGAN2 and a HyperInverter encoder, the original face image ($x_{org}$) is mapped into a latent space.
- Semantic Editing: Users define acceptable ranges for facial attribute changes (e.g., face roundness, eye openness). The system identifies "edit vectors" within this latent space that correspond to these attributes.
- Optimization via PGD: To ensure maximum privacy, the system employs Projected Gradient Descent (PGD) to optimize the edit vector. The objective is to maximize the Identity Loss ($L_{id}$) between the original face and the generated scapegoat ($\tilde{x}$), ensuring the identity similarity is minimized while staying within the user-defined "style neighborhood."
Experimental Results
The authors evaluated the method using the CelebA-HQ dataset and the ArcFace recognition model.
Identity Masking and User Perception
Quantitative analysis showed a dramatic decrease in identity similarity as the editing intensity ($\epsilon$) increased. At an intensity of $\epsilon=0.05$, the similarity score dropped from a baseline of 0.855 to 0.203. A user study corroborated these findings, with participants rating scapegoat images as recognizable avatars that effectively masked the specific identity of the original subject.
Deepfake Resilience
A critical advantage of Scapegoat Generation is its resilience to subsequent deepfake synthesis. When a scapegoat image is used as a target for a deepfake model (such as SimSwap), the resulting "scapegoat deepfake" ($\tilde{y}$) inherits the identity mismatch. Because the protection is embedded in the semantic features of the image rather than high-frequency noise, it remains effective regardless of the deepfake architecture used by an attacker.
Conclusion and Future Work
Scapegoat Generation offers a user-centric, model-agnostic alternative to traditional deepfake defenses. By allowing users to curate their digital representation, it provides a robust layer of privacy that persists through the synthesis process. Future work aims to expand the range of available semantic edits and integrate text-to-edit capabilities to enhance user control and accessibility.
Summary updated