Cross-Modal Residual Cross-Attention for Multimodal Bot Detection
A cross-modal fusion module (CMRCA) that integrates heterogeneous social-media account features from separate visual (profile image) and textual (user metadata, tweet text) encoders using cross-attention between modalities combined with residual connections, feeding a classifier that outputs an account's bot probability. It replaces naive concatenation of multimodal embeddings and forms the core of the one-stage end-to-end MSM-BD detector, validated on TwiBot-22.
2501.00204
MSM-BD is a one-stage, end-to-end multimodal social-media bot-detection model for platform X. It ingests heterogeneous per-account information across two modalities: visual (profile image) and textua…