Designing Accessible Obfuscation Support for Blind Individuals’ Visual Privacy Management

Document Title

Designing Accessible Obfuscation Support for Blind Individuals’ Visual Privacy Management

Document Information

• Subject Area: Human-Computer Interaction and Accessibility Design, Visual Privacy Management
• Keywords: Accessibility Technology, Privacy Protection, Blind Photography, Obfuscation Design, Artificial Intelligence, User Research, Computer Vision, User Control, Human-Computer Interaction

Research Background and Issues

• What problems or challenges did the authors identify?

  • Blind individuals often encounter visual privacy issues when taking photos, including assessing and removing potentially sensitive information from images.
  • Current visual privacy tools are typically designed for sighted users and fail to meet the needs of blind individuals for independent privacy management.
  • Blind individuals desire control over private content, but existing tools lack sufficient design and interaction support for this requirement.

• Why is this issue important?

  • Privacy protection is a critical societal concern, particularly challenging for blind individuals who rely on third-party services.
  • Improving visual privacy management tools can enhance user independence and security while promoting social inclusion and technological equity.

• Research Motivation and Related Work

  • Building on prior research into blind individuals’ privacy management in photography, the authors aim to extend existing designs to create independently operable obfuscation tools.
  • Previous studies have shown that blind individuals lack information and control over privacy management tools, expressing a desire for flexible and adjustable solutions.

Solution

• What methods or solutions did the authors propose?

  • Developed a screen reader-accessible, mid-fidelity obfuscation tool prototype and explored design scenarios through user research.
  • Created two versions of the tool prototype: one using existing AI models (e.g., SAM, BLIP2, ChatGPT) for automatic content generation, and another manually optimized by researchers (Wizard-of-Oz) to simulate the ideal tool’s functionality.

• What is innovative about this solution?

  • Investigated blind users’ mental models and interaction challenges with AI-assisted privacy obfuscation technology.
  • Proposed design guidelines for visual privacy obfuscation tailored to blind users, emphasizing non-visual interaction, user control, and improved descriptions of obfuscation effects.
  • Conducted user research to analyze the impact of inaccuracies in AI models on user experience when detecting and processing private visual content.

• What implementation steps were taken, and what key technologies were used?

  • Utilized AI models (e.g., SAM segmentation model, BLIP2 image description model) to locate objects in images and employed ChatGPT to detect and classify potential privacy-sensitive objects.
  • Provided three common obfuscation methods: blurring, blacking out, and background filling (implemented via LaMa tool).
  • Designed a user interaction interface, including image navigation and editing features, supported by touch and screen reader-based non-visual interaction.
  • Conducted user research involving 12 blind participants to explore usage patterns, feedback, and opportunities for design optimization.

Research Outcomes

• What specific results were achieved?

  • Blind users demonstrated strong learning ability regarding visual concepts (e.g., foreground and background, obfuscation effects) but faced challenges in understanding specific obfuscation results.
  • Participants expressed diverse creative control needs, including personalized settings for obfuscation styles and selection ranges.
  • Proposed design recommendations to reduce random errors and enhance the quality of obfuscation effect descriptions.

• How does it compare to existing solutions?

  • Places greater emphasis on the specific needs and interaction experiences of blind users, highlighting user autonomy and tool flexibility.
  • Improved the application of existing models, making the detection of private objects more precise and suitable for blind users.

• What were the experimental or evaluation results?

  • The study revealed limitations in AI models for detecting and describing obfuscation effects, such as misclassification of specific visual effects (e.g., blurred areas).
  • Blind participants were more satisfied with the Wizard-of-Oz results but expressed concerns over deviations in current models, offering multiple suggestions for improvement.

• Limitations and Future Directions

  • Limitations: The prototype design remains at a mid-fidelity stage, lacking exploration of multi-object obfuscation handling and comprehensive usage scenarios; existing technology has insufficient computational capacity for real-time obfuscation processing.
  • Future Directions: Develop high-fidelity tool prototypes to achieve more accurate visual content detection and obfuscation techniques; design tools supporting more complex collaborative needs, such as joint editing by blind and sighted users; extend the tool to general image editing domains to broadly support blind individuals’ digital content creation.

This study provides new insights into designing visual privacy management technologies for blind individuals, contributing positively to future research in the fields of accessibility and human-computer interaction.