cs.CVMar 29, 2026

AI-Powered Facial Mask Removal Is Not Suitable For Identification

Authors: Emily A Cooper, Hany Farid

Organizations: Herbert Wertheim School of Optometry & Vision Science University of California, Berkeley · School of Information University of California, Berkeley

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Recently, crowd-sourced online criminal investigations have used generative-AI to enhance low-quality visual evidence. In one high-profile case, social-media users circulated an "AI-unmasked" image of a federal agent involved in a fatal shooting, fueling a wide-spread misidentification. In response to this and similar incidents, we conducted a large-scale analysis evaluating the efficacy and risks of commercial AI-powered facial unmasking, specifically assessing whether the resulting faces can be reliably matched to true identities.

Figures & tables

Figure 1: A representative set of original unmasked faces (top) and the result of unmasking using Gemini to fill in the removal of the lower-half of the face. The value above each face is the biometric similarity (in [−1,1][-1,1] ) between the original and unmasked face. Larger values correspond to a higher facial similarity.
Figure 2: Senator Elizabeth Warren’s mask (left) is removed (middle) using ChatGPT (photo sources: commons.wikimedia.org).
Figure 3: Distributions of biometric facial similarity scores: (A) pairs of images of the same identity (green), images of different identities from the same racial/gender group (red), and images of Doppelgangers (yellow); (B)-(D) pairs of a reference and unmasked image in which the lower-half was removed; and (E) pairs of a reference image and an image in which COVID masks were removed.
 non matchnon match
false matchSame-IDAI-unmaskedSame-IDAI-unmasked
1.0%0.0000%8.1%0.0000%6.3%
0.1%0.0000%17.3%0.0000%13.7%
0.01%0.0001%28.6%0.0000%23.1%
0.001%0.0006%40.7%0.0001%33.8%
0.0001%0.0049%52.3%0.0012%44.5%
Table 1: Face detection accuracy for a set of fixed false match rates (incorrectly identifying different identities as same). For each false match rate, non match rates are reported: the percentage of image pairs of the same identity incorrectly identified as different. The “Same-ID” column corresponds to the Same-ID dataset, and the “AI-unmasked” column corresponds to the combined ChatGPT and Gemini unmasking of the Senators dataset (see Figure 3 (A) and (E)).
More in this topic

Digital Image Forensics

Digital image forensics develops methods to detect, localize, explain, and trace image manipulation, synthetic content, and hidden data.

See all 112 papers

Explore similar work

CardsList

  1. FACT: A Forensic Agent with Compiled Tool-Use Trajectories for AI-Generated Image Detection

    Sep 5, 2026Jiaoyang Chen, Bin Hu, Jingyu Hu +3AI-Generated Image DetectionImage Forgery Detection

    Lingnan University, Hong Kong · Harbin Institute of Technology, Shenzhen · The University of Hong Kong +2

  2. IdentiFace: Multi-Modal Iterative Diffusion Framework for Identifiable Suspect Face Generation in Crime Investigations

    May 1, 2026Weichen Liu, Yixin Yang, Changsheng Chen +1Diffusion SamplingImage Generation

    Southeast University · Nanjing, China · Shenzhen MSU-BIT University +1

  3. EEG and Eye-Tracking Evidence That AI Disclosure Shapes Face Evaluation

    Oct 7, 2026Teodora Mitrevska, Luise Donat, Andreas Butz +3

    LMU Munich, Munich, Germany · HU Berlin, Berlin, Germany · Centrum Wiskunde & Informatica, Amsterdam, Netherlands +2