Scam Shield: Multi-Model Voting and Fine-Tuned LLMs Against Adversarial Attacks