Phishing attacks continue to pose a significant cybersecurity threat by exploiting users through deceptive emails. While traditional centralized phishing detection methods achieve high accuracy, they face critical limitations, including privacy concerns, scalability challenges, and susceptibility to data distribution shifts. To address these issues, MBFT-Lite (Multi-Branch Federated Transformer-Lite) is proposed—a novel, lightweight, privacy-preserving phishing detection framework based on federated learning. MBFT-Lite employs a dual-branch architecture that combines semantic analysis via a TinyBERT model with structural analysis through XGBoost models on metadata features. The TinyBERT models are aggregated using the FedAvg a lgorithm, while the XGBoost models remain decentralized to enhance privacy and reduce communication overhead. Predictions from both branches are fused using strategies such as weighted averaging, max-confidence voting, and a meta-classifier. Extensive evaluations across four diverse email datasets—CEAS 08, SpamAssassin, Nigerian Fraud Emails, and Nazario—demonstrate that MBFT-Lite achieves over 99.4% accuracy across clients and 97.9% accuracy on unseen data, while preserving data privacy and ensuring scalability.