Blockchain-based personalized federated learning framework for drug recommendation systems resilient to model poisoning

Abstract Federated learning enables multiple healthcare entities to collaboratively train a global model while ensuring patient data privacy through local model training without sharing raw data. However, FL remains vulnerable to adversarial attacks such as model poisoning, data injection, and model inversion that compromise model integrity. To address these challenges, this paper presents a blockchain-based personalized federated learning (FL) framework designed to enhance the security, privacy, and efficiency of decentralized model training in healthcare environments. It integrates Practical Byzantine Fault Tolerance (PBFT) for tamper-resistant aggregation, L2-norm anomaly filtering for lightweight adversarial defense, and a Neural Architecture Search (NAS)-optimized hybrid Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) model with attention to enable efficient, personalized modeling of non-IID healthcare data. Together, these components address key FL challenges, including robustness to model poisoning, accuracy, and deployment on resource-constrained devices. To evaluate its effectiveness, we apply the proposed framework to drug recommendation tasks using three real-world medical datasets, namely Symptom2Disease, UCL Drug, and Dermo Questions, achieving F1-scores of 0.97, 0.70, and 0.83, respectively. The framework demonstrates competitive performance compared to conventional and state-of-the-art methods while significantly reducing the number of trainable parameters, highlighting its suitability for real-time, on-device healthcare applications. These results validate the framework’s ability to deliver secure, personalized, and privacy-preserving recommendations in intelligent healthcare systems.