Joint Semantic Detection and Dissemination Control of Phishing Attacks on Social Media Via Llamabased Modeling

This paper tackles the growing threat of phishing attacks on social media by proposing a unified framework that combines content detection with propagation control, built upon the LLama large language model. The approach begins by extracting deep semantic representations from social media text using LLama, followed by adversarial training to enhance robustness and classification accuracy against deceptive messages. To address the spread of phishing content, a social network propagation graph is constructed, and the diffusion process is simulated using an influence-based model. A targeted intervention strategy is then introduced to disrupt high-risk propagation paths. To achieve a balanced performance, the framework employs a joint optimization objective that integrates semantic classification loss, adversarial loss, and propagation control loss. Experiments on real-world Twitter data show that the proposed method consistently outperforms existing approaches across multiple evaluation metrics, demonstrating strong detection capabilities and effective intervention.