Cross-Modal Anomaly-Bridged Instance and Feature Contrastive Learning for Social Bot Detection

The rapid proliferation of social bots threatens the reliability of online platforms. Existing detection methods relying on single-modality or weak multimodal fusion often fail to generalize across diverse behaviors. We propose CABINET—a cross-modal anomaly-bridged instance and feature contrastive learning framework for robust social bot detection. CABINET integrates textual and structural information via two key modules: anomaly-targeted text augmentation (ATTA), which enhances semantic discriminability by emphasizing anomalous cues, and hierarchical contrastive learning (HCL), which aligns embeddings across instance and feature levels to maintain cross-modal consistency. This design effectively bridges semantic–structural gaps and improves robustness to noisy data. Experiments on Cresci-15 and TwiBot-20 show that CABINET outperforms state-of-the-art methods, achieving F1 scores of up to 99.40% and 90.64%, respectively. Ablation and sensitivity analyses further confirm the effectiveness and stability of each component

Feature extraction · Feature learning · Semantic feature · Anomaly Detection Techniques and Applications · Misinformation and Its Impacts · Spam and Phishing Detection