BotCHF: camouflage-heterogeneity-aware fusion for social bot detection
TL;DR
A camouflage-heterogeneity-aware decision fusion framework for social bot detection (BotCHF), which adopts an alternating collaborative optimization strategy that periodically injects fine-tuned semantic features into the graph encoder, thereby progressively aligning semantic encoding with structural learning.
Abstract
Social bot detection is crucial for maintaining the security and integrity of online social networks (OSNs). Although graph-based methods have achieved state-of-the-art performance, rapid advances in large language models have made bots increasingly similar to humans in the textual modality, while some bots also adopt diverse camouflage strategies in the structural modality. As a result, social bots exhibit pronounced individual-level heterogeneity in camouflage behavior, causing the discriminative power of different modalities to vary substantially across accounts. Existing multimodal methods typically rely on unified fusion strategies, which are insufficient to handle such sample-specific variation and may lead to misclassification when one modality is heavily camouflaged. Moreover, they generally lack an effective mechanism for jointly optimizing semantic and structural representations. To address these issues, we propose a camouflage-heterogeneity-aware decision fusion framework for social bot detection (BotCHF). In the encoding stage, BotCHF adopts an alternating collaborative optimization strategy that periodically injects fine-tuned semantic features into the graph encoder, thereby progressively aligning semantic encoding with structural learning. In the fusion stage, it maintains separate text and graph branches and adaptively weights their predictions for each account, enabling decision fusion that responds to account-specific variation in modality reliability. Extensive experiments on three real-world datasets demonstrate that BotCHF consistently outperforms strong baselines. Further analysis of fusion weights reveals substantial cross-dataset differences in modality preference, highlighting the necessity of explicitly modeling camouflage heterogeneity for robust social bot detection.