Are You Comfortable Sharing It?: Leveraging Image Obfuscation Techniques to Enhance Sharing Privacy for Blind and Visually Impaired Users
Satabdi Das, Nahian Beente Firuj, Manjot Singh, Arshad Nasser, Khalad Hasan · 2026 · Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI '26) · doi:10.1145/3772318.3791916
Summary
This CHI 2026 paper addresses a privacy gap in how blind and visually impaired (BVI) users share photos: because they cannot visually inspect what they capture, images routinely contain sensitive or inappropriate content (identifiable people, nudity, documents with credentials, medical details, revealing backgrounds) that the sharer is unaware of. The authors argue that existing vision-based safety classifiers are designed for sighted users and do not expose controls that fit non-visual workflows, and that prior BVI privacy work has largely focused on single obfuscation techniques or narrow content categories. Grounding the work in Nissenbaum's contextual integrity and Privacy by Design principles, the paper systematically varies content category, obfuscation method, and recipient audience to understand how these factors jointly shape BVI users' comfort. The study ran with 20 BVI participants (ages 20-31, mean 24.15; 13 totally blind or totally blind from birth, 7 with limited light perception) recruited via community networks. Participants saw 20 images across 10 sensitive-content categories (identity of people, children, nudity, private areas at home, embarrassing shots, identity information, medical condition, habit/interest, illegal/inappropriate content, personal moments) and evaluated four obfuscation techniques: blurring, pixelation, masking, and content-based filling (AI-generated replacement). Each technique was described to participants with non-visual metaphors (e.g., pixelation as "a misty or cloudy image", masking as "the beep used to hide private information in phone call recordings"). Comfort was measured on a 7-point Likert scale across three target audiences — family, friends, strangers — both before and after filtering. Analysis used Friedman tests with Bonferroni-corrected Wilcoxon post-hoc.
Key findings
Pixelation was consistently the least preferred technique across categories, particularly for nudity (C3, p=.001), private areas (C4, p=.003), identity information (C6, p=.001), medical conditions (C7, p=.004), and illegal content (C9, p=.004). Participants rejected pixelation as "blocky", aesthetically disruptive, and too coarse to preserve useful context for the sighted recipient who might still need to interpret the image. Blurring was preferred for highly sensitive content where participants wanted the sensitive element removed but the composition preserved; masking and content-based filling were preferred when explicit details needed to be fully replaced. No single technique was universally preferred — preference was strongly content- and context-dependent. Before filtering, participants were significantly more comfortable sharing with family and friends than with strangers across all 10 categories. After filtering, comfort rose significantly for every category-audience pair; the largest absolute gains were in the stranger condition, where filtering partially restored contextual integrity. Even after filtering, residual discomfort remained for nudity and identity information when shared with strangers (medians 3.50 and 3.00 respectively), indicating that obfuscation cannot fully override deeply personal concerns. Qualitative comments highlighted reliance on AI summaries and visual-description services: coarse pixelation degrades those downstream descriptions, so participants treated it as a last-resort default. The paper proposes three design directions: prioritise context-preserving filters over pixelation, offer reversible and persistent user-driven personalisation rules (e.g., "auto-blur identity information"), and implement intelligent content- and audience-aware filtering that adapts obfuscation strength to the intended recipient.
Relevance
For practitioners building VDS, visual-assistance, or any image-sharing feature used by BVI users, the paper is an unusually specific evidence base: it gives both rank-ordered technique preferences by content category and measurable audience-sensitivity deltas. Actionable implications include letting users set persistent, reversible auto-obfuscation rules rather than forcing per-share decisions; surfacing obfuscation choices via non-visual previews (descriptive text of what will remain visible); and tuning default strength by recipient identity. The work also gives product teams a vocabulary for why pixelation — the default in many mainstream social and messaging platforms — is the wrong default for BVI workflows. Limitations to flag: 10 categories with only 2 images each, all participants from a single metropolitan area, no measurement of downstream utility (can the recipient still interpret filtered images? can AI captioning still describe them accurately?), and no measurement of whether participants already use any filtering strategies today. The four obfuscation techniques are well-known but were described to participants via metaphor only — comprehension effects may partially explain pixelation's poor showing. The study measures comfort and willingness-to-share, not real sharing behaviour.
Tags: blind and low vision · privacy · image obfuscation · visual description services · user study · image sharing