Abstract
Gaussian blur is widely used to anonymize faces in sensitive imagery, under the assumption that it irreversibly destroys identity-defining information. This paper introduces Revelio, a multi-stage deanonymization framework that utilizes conditional diffusion models and identity retrieval to reconstruct faces from heavy blur. Experimental results demonstrate that Revelio achieves a 95.9% re-identification accuracy on the CelebA-HQ dataset, significantly outperforming existing blind image restoration methods. These findings suggest that Gaussian blur is fundamentally insufficient for privacy protection and should be replaced by hard redaction techniques.
Introduction
In contexts such as political protests or crime scene reporting, individuals often apply Gaussian blur to photographs to protect the identities of those pictured. While a blurred face may appear as a featureless blob to the human eye, the mathematical structure of the original image is often preserved in a latent state. Existing restoration tools typically focus on general image quality (e.g., removing camera shake) rather than identity preservation. This study investigates whether an adversary can specifically recover the original identity of a blurred subject, even when facial features are completely obscured.
The Revelio Framework
The authors propose Revelio, a pipeline designed for identity-aware restoration:
- Kernel Estimation: An EfficientNetV2-based regression model estimates the Gaussian kernel size used for the blur.
- Base Restoration ($M_B$): A conditional diffusion model, trained to invert Gaussian blur, produces a structural approximation of the face.
- Identity Retrieval ($M_R$): The system queries a reference database (e.g., CelebA-HQ) to find the most likely identity match for the preliminary restoration.
- Fidelity Enhancement ($M_F$): The diffusion model is fine-tuned using clear images of the retrieved identity, allowing it to "hallucinate" person-specific details like eye shape and skin texture.
Experimental Results
Revelio was evaluated against several state-of-the-art (SOTA) baselines, including GFP-GAN, DifFace, and PULSE. Under "heavy blur" conditions (kernel size 81), Revelio achieved a 95.9% identity retrieval accuracy.
As shown in the figure above, while baseline models often produce high-quality faces, they frequently fail to preserve the specific identity of the original subject. Revelio's identity-aware approach ensures that the reconstructed image is not just a realistic face, but the correct face.
Discussion and Security Implications
The success of Revelio highlights a "memorization effect" in generative models, where the system uses its training data to fill in missing information. This poses a significant threat even for "Out-of-Distribution" (OOD) scenarios; the system can identify unseen faces with 90% accuracy, preventing false matches. Furthermore, the attack is robust against common defenses like noise addition or JPEG compression when the model is adaptively retrained.
Conclusion
The research concludes that Gaussian blur provides a false sense of security. For effective anonymization, the authors recommend "hard" redaction—completely covering sensitive areas with solid, opaque blocks—to ensure that no recoverable information remains in the image file.
Summary updated