ReCAST: Restoration-aware Cascaded Stage-wise Training for Obfuscated SMS Risk Classification

The researchers proposed the ReCAST framework, aimed at addressing the difficulty of deploying large models in anti-interference tasks. This framework uses supervised learning for deconfusion span detection, deconfusion type prediction, and text restoration, and distills the deconfusion capabilities of large models into smaller student models, enabling their use in downstream risk classification tasks. Experiments were conducted on a real-world Chinese SMS benchmark built internally, and the results showed that this method significantly improved anti-interference classification performance compared to direct training of baseline models. The study believes that this perceptual restoration distillation approach provides a feasible path for deploying small-scale, highly robust SMS risk classification systems in production environments.