WSMS: weakly supervised multi-feature steganalysis with EVT calibration for cloud-edge collaborative intelligence

Abstract The rapid growth of large language models introduces new security risks through steganographic text that closely mimics natural language. To address this challenge in cloud–edge AI ecosystems, we propose WSMS (Weakly Supervised Multi-feature Steganalysis), which integrates statistical, structural, semantic, and logical features via adaptive fusion. WSMS employs weakly supervised self-training with EVT-based threshold calibration, confidence-weighted pseudo-labeling, and teacher–student consistency. Experiments across multiple datasets show that WSMS effectively detects LLM-generated steganographic content and generalizes well under weak supervision. Notably, WSMS achieves 2% higher accuracy in few-shot scenarios and 8% higher accuracy under imbalanced conditions compared with baselines, demonstrating its scalability and reliability for cloud–edge collaborative AI security.