Researchers Show How AI Can Reconstruct Blurred and Pixelated Faces

Blurring and pixelation, techniques used for decades to hide the faces of witnesses, whistleblowers, and protesters in video footage and photographs, are no longer reliable protection for anonymity. A team of researchers at the University of Illinois Urbana-Champaign has developed a method called Revelio that uses a diffusion model to reverse Gaussian blur and identify the hidden identity underneath with very high accuracy.

Gaussian blur and pixelation have long been treated as a standard way to protect the identity of people recorded without their consent or who request anonymity, such as whistleblowers, crime victims, or protesters. The problem is that mathematically, both processes are a predictable form of image degradation, and that is exactly the kind of distortion diffusion models handle best, the same class of neural networks that powers image generators like Midjourney and Stable Diffusion.

How Revelio Works

The system described by the Illinois team uses a conditional diffusion model combined with an identity search against a reference database. The process runs in three stages: first a preliminary facial reconstruction is generated, then the algorithm compares it against a database of known faces to find the most likely match, and finally it refines the result by fine-tuning the model on other available photos of the same person. The key mechanism is the memorization effect in generative models trained on millions of faces, which lets them reconstruct features even from a heavily degraded image.

Reconstructing faces in video remains harder than in single photos, since it requires keeping features consistent across consecutive frames. The researchers stress, however, that the final result is not always a faithful copy of the real face, but rather the statistically most likely version matching the degraded material, which in practice can still be enough to form an accurate hypothesis about someone's identity.

A Threat to Journalism and Human Rights

The consequences fall mainly on organizations that document human rights abuses and on investigative journalism. WITNESS, which has advised on safe video documentation practices for years, and Human Rights Watch have already changed how they film people who need their identities protected. Amnesty International previously faced criticism for using generative AI to edit evidentiary footage, which shows how thin the line is between protecting a subject and manipulating the material.

Nothing is fully resistant to future technology - Hany Farid, media forensics expert

In January 2026, NPR reported on a case in which internet users tried to identify a US Immigration and Customs Enforcement (ICE) officer from footage with a blurred face, using exactly this kind of AI tool. Similar techniques could theoretically be used to identify witnesses in criminal cases or unmask protesters in countries with repressive governments.

What Comes Next for Anonymization

Experts say the only real alternative to blurring is irreversible methods, such as fully replacing a face with a synthetic avatar, masking it with a black box, or removing sensitive parts of the image entirely rather than degrading them. The US Department of Justice, for instance, used black redaction boxes when it released photos related to the Epstein case around late January 2025, which some specialists point to as a more robust solution than traditional blurring.

For Polish newsrooms, NGOs, and institutions that publish material requiring the anonymization of witnesses or victims, the findings mean existing standards need to be revisited. Blurring a face alone is no longer sufficient protection, especially in an era of widely available, increasingly cheap generative tools capable of automatically reconstructing images.

Sources: Tech Policy Press (techpolicy.press), "Restoring Gaussian Blurred Face Images for Deanonymization Attacks" (arxiv.org)