PhishGuard implements a unified framework combining LLaMA-powered semantic detection with graph-based propagation control for identifying and mitigating phishing attacks on social media. The system uses a LLaMA encoder with deep semantic embeddings and attention mechanisms, fine-tuned via LoRA for phishing-specific risk assessment. It incorporates adversarial training using KL divergence for robustness and combines social network graph analysis with temporal patterns to predict and intervene in attack dissemination. The architecture includes a 2-way classifier head with confidence scoring for phishing detection alongside propagation modeling for targeted intervention.