FediScan: Collaborative Social Bot Detection in the Fediverse

The growing concern for data privacy and user autonomy has led to the rise of decentralized online social networks, such as Mastodon. Unlike centralized platforms, Mastodon's federated architecture comprises a number of independent instances. Social bots, which are automated accounts that might spread misinformation and manipulate discourse, pose significant threats to platform moderation and security. Detecting these social bots in decentralized online social networks such as Mastodon is challenging due to the fragmented governance, non-IID data distributions, and diverse modalities across their different instances. Current social bot detection methods, designed for centralized systems, fail to address these challenges while preserving user privacy. To fill this gap, we propose FediScan, a decentralized federated learning framework for social bot detection in the Fediverse. FediScan introduces three key innovations: (1) a modality-specific data augmentation module integrating a feature augmentation strategy and a multimodal encoder with a gated attention mechanism to learn informative user representations for robust social bot detection; (2) a semantic-aware communication protocol incorporating an instance hypergraph built upon hashtag co-occurrence, enabling knowledge sharing without exchanging raw data; and (3) an asynchronous aggregation strategy to accelerate convergence and reduce overhead. Extensive evaluation on a representative multimodal dataset from Mastodon demonstrates that FediScan achieves a significant improvement in F1-score over existing methods. This work introduces a novel approach for privacy-preserving, collaborative detection of social bots within decentralized online social networks.