Detecting and Defending Against Seizure-Inducing GIFs in Social Media

Document Title

Detecting and Defending Against Seizure-Inducing GIFs in Social Media

Document Information

• Subject Area: Human-Computer Interaction, Cybersecurity, Accessibility Research
• Conference: CHI 2021
• Keywords: Accessibility, Photosensitive Epilepsy, GIF, Human-Computer Interaction, Cybersecurity, Consumer Protection, Browser Extension, Visualization

Research Background and Problem

• Issues and Challenges:
  • Photosensitive Epilepsy (PSE) is a chronic condition triggered by specific light stimuli.
  • Flashing or strobing GIFs (rapidly flickering animated images) can cause severe physical harm to individuals with photosensitive epilepsy, potentially leading to life-threatening situations.
  • Current creator-driven protection systems (e.g., PEAT and Harding FPA) are ineffective against malicious attacks and fail to provide adequate defense.
  • Malicious actors have exploited social media to disseminate seizure-inducing GIFs, deliberately designing attacks that have caused real harm.
• Significance:
  • Multiple cases since 2016 have demonstrated that malicious dissemination of seizure-inducing GIFs on social media platforms is becoming easier, posing significant risks to individuals with epilepsy.
  • Existing web accessibility guidelines and tools show clear limitations in addressing this issue.
• Research Motivation and Related Work:
  • Society has insufficient awareness of photosensitive epilepsy, and online protective measures are weak.
  • Existing browser extensions (e.g., EpilepsyBlocker) have limited detection accuracy and effectiveness, with insufficient experimental evaluation.
  • Current standards do not account for GIF looping behaviors or hazardous patterns in repetitive designs, and systematic research in this area is lacking.

Solution

• Method or Solution:
  • Proposed and designed a browser extension named "PhotosensitivityPal" to detect and mitigate potential seizure-inducing GIFs and videos for consumers.
  • Developed a framework categorizing defense systems into three types: creator-driven, platform-driven, and consumer-driven.
  • Defined four key design requirements for consumer-end systems based on user studies and reviews of existing systems.
• Innovations:
  • First to propose an integrated detection system combining photosensitivity risk factors, including flashing, saturated red transitions, and repetitive patterns.
  • Introduced user-controlled mitigation strategies for hazardous content (e.g., low contrast and grayscale filters).
  • Created standardized evaluation datasets to validate system performance.
• Implementation Steps and Key Technologies:
  1. Core Algorithm Implementation:
     • Detection of flashing and red transitions based on WCAG 2.0 standards.
     • Repetitive pattern detection using empirical modeling algorithms (e.g., linear stripes, dot patterns).
  2. Browser Extension Design:
     • Automated detection, alert, and blocking functionalities.
     • Mitigation of potential hazardous content through low contrast and desaturation options.
  3. Experiments and Evaluation:
     • Validated system performance using simulated GIFs, random GIFs, and manually collected potentially hazardous GIF datasets.

Research Outcomes

• Specific Outcomes:
  • Tool Development: Developed PhotosensitivityPal, capable of detecting and mitigating photosensitivity risk content.
  • Dataset Contribution: Created three standardized datasets, including 150 simulated GIFs, 200 randomly collected GIFs from social media, and 137 manually annotated potentially hazardous GIFs.
  • Risk Assessment Metrics:
    • Detected at least one photosensitivity risk factor in 8.15% of social media GIFs.
    • The most common risk factor was flashing (4.70%), followed by repetitive patterns (3.35%), and saturated red transitions (0.6%).
• Comparison with Existing Solutions:
  • Compared to PEAT, PhotosensitivityPal demonstrated improved accuracy, recall, and precision on simulated datasets, with a lower miss rate.
  • Outperformed EpilepsyBlocker by functioning effectively across more diverse datasets.
• Experimental Results:
  • Social media platforms provide insufficient protection for photosensitive users, with 8% of analyzed GIFs containing hazardous content.
  • The risks posed by video looping behaviors are underestimated: many maliciously designed looping GIFs were not correctly flagged by existing detection systems.
• Limitations and Future Directions:
  • Limitations:
    • Detection capabilities for repetitive patterns are constrained, potentially missing complex high-risk designs.
    • Data sources are limited, and manually collected hazardous content may introduce bias.
  • Future Research:
    • Update empirical standards for photosensitive epilepsy trigger risks and develop additional testing standards for mobile devices.
    • Develop advanced detection algorithms based on machine learning.
    • Explore automated methods for distinguishing malicious attacks from accidental triggers using contextual understanding.

Conclusion

This research developed a novel consumer-driven browser extension, PhotosensitivityPal, significantly enhancing online protection for individuals with photosensitive epilepsy. It provides valuable insights into cybersecurity and accessibility for social media and the broader technology community.