Naufal El Khatibi
Please Note
1 records found
1
Visual social media data provides unprecedented opportunities for studying riot dynamics at fine spatiotemporal scales, yet analyzing such content raises significant privacy concerns as images can identify individuals participating in politically sensitive events. This creates tension between analytical transparency and ethical obligations that existing methods inadequately address. We propose a privacy-preserving methodology for identifying riot-related content from visual social media posts by converting images to text descriptions during data collection, then classifying these captions rather than analyzing raw visual content. Using France’s 2023 riots as a case study, we collect 107,674 Snapchat posts (19% of images, 81 % of videos) and process them through a BLIP2 image-to-text model, generating descriptive captions at a frame rate of 2 frames per second. A BERT classifier trained on 2,084 manually annotated captions achieves an F1 score of 0.73±0.03 at video-level classification, approaching the performance of direct computer vision methods (F1=0.72) while substantially enhancing privacy protection. The methodology enables researchers to share interpretable text data for replication and secondary analysis, reconciling open science principles with privacy requirements. While computational costs increase significantly (2.7B vs 5M parameters), this trade-off represents a meaningful advance for ethically studying social movements through visual social media data.