Adaptive Privacy Preserving Federated Learning for LLM Based Healthcare IoT Systems

This work evaluates a privacy-preserving federated learning framework for adapting large language models to healthcare text without centralizing raw data. We simulate healthcare-IoT gateways using public healthcare text datasets, including med-dialogue-1M, PubMedQA, and MTS-Dialog, and fine-tune Qwen2.5-1.5B-Instruct using rank-4 LoRA updates. The system combines federated optimization with adaptive differential privacy, CKKS-based encrypted aggregation, and ECDSA signature checks to study privacy, confidentiality, and update integrity in a reproducible simulation setting. Across staged experiments, we compare FedAvg, FedProx, and FedYogi under matched privacy budgets, gateway sampling, dropout conditions, and repeated random seeds. Final results show that FedYogi with adaptive DP achieves lower validation perplexity than fixed-DP FedYogi at the same privacy budget, while FedAvg shows almost no adaptive gain. These findings suggest that adaptive privacy control interacts meaningfully with federated optimizer choice for privacy-preserving LLM adaptation. The study does not claim clinical readiness or real gateway deployment, but provides a reproducible experimental basis for future secure healthcare-IoT LLM research.