American Sign Language Video Anonymization to Support Online Participation of Deaf and Hard of Hearing Users
Sooyeon Lee, Abraham Glasser, Becca Dingman, Zhaoyang Xia, Dimitris Metaxas, Carol Neidle, Matt Huenerfauth · 2021 · ASSETS '21: The 23rd International ACM SIGACCESS Conference on Computers and Accessibility · doi:10.1145/3441852.3471200
Summary
Deaf and Hard of Hearing (DHH) individuals who prefer to communicate in American Sign Language (ASL) face a unique privacy barrier online: because ASL is a visual language conveyed through movements of the face, head, hands, and torso, and because the face carries essential linguistic information (grammatical markers via eyebrow movements, mouth morphemes, head tilts), signers cannot simply obscure their face to remain anonymous the way speakers of written or spoken languages can. This means DHH ASL users are effectively unable to discuss sensitive topics — health concerns, legal questions, political views, or personal experiences — anonymously in their primary language. Lee et al. evaluated emerging face-transformation technologies as a potential solution, testing three prototype approaches with 16 DHH ASL signers in 70-minute interview sessions conducted via Zoom. The prototypes were: (1) "with-torso," which replaces the signer's face with a synthetically generated face from a target photograph while preserving facial expressions, head movements, and the original torso and background; (2) "without-torso," identical to with-torso but replacing the torso and background with a flat gray color for additional anonymity; and (3) "tiger-face," a simpler filter that overlays an animated cartoon tiger head on the signer's face without preserving facial expressions. The three-phase study had participants attempt to identify disguised signers from photo line-ups, rate transformed videos on Likert scales for understandability, naturalness, and anonymity protection, and then view transformed videos of themselves to evaluate personal acceptability.
Key findings
The study revealed a fundamental three-way trade-off among understandability, naturalness, and anonymity. With-torso videos were rated most understandable (81% strongly agreed) and most natural in appearance, but provided weaker anonymity because the signer's body, clothing, and background remained visible — 9 of 16 participants correctly identified the male signer from a photo line-up. Without-torso videos offered the strongest anonymity protection (88% agreed identity was disguised) while maintaining reasonable understandability (62% strongly agreed), but were perceived as less natural due to the gray background. Tiger-face videos were least understandable (only 25% strongly agreed) because they eliminated all facial expressions, which are grammatically essential in ASL. Critically, half of participants shifted their prototype preference from with-torso to without-torso after seeing their own transformed videos, because they noticed recognizable personal details (clothing, jewelry, body shape) in the with-torso versions. Participants overwhelmingly wanted transformation characteristics — particularly race, age, and hair style — to closely match their own appearance for naturalness. Nearly all participants saw value in this technology for discussing sensitive topics online, though they raised ethical concerns about impersonation and misuse, suggesting safeguards such as disclaimers indicating a video has been transformed and using computer-generated faces rather than real people's photographs.
Relevance
This research addresses a significant but often overlooked equity gap in online participation: while hearing individuals can easily post anonymous text, DHH ASL users are forced to choose between revealing their identity or communicating in a language that is not their primary one. The findings have direct implications for designers of video-based communication platforms, social media applications, and any system that aims to support anonymous participation. The identified trade-offs between understandability, naturalness, and anonymity provide a practical framework for designers building face-transformation tools for sign language video. The study also raises important considerations around linguistic accessibility — the fact that facial expressions in ASL carry grammatical meaning (not just emotional expression) means that any privacy-preserving technology must be designed with linguistic expertise to avoid destroying the message itself. For accessibility practitioners, this work highlights that privacy and anonymity are accessibility concerns that intersect with language modality in ways that standard web accessibility guidelines do not currently address.
Tags: deaf and hard of hearing · sign language · American Sign Language · anonymization · privacy · face transformation · online participation · video communication