Mesopotamian Journal of Artificial Intelligence in Healthcare
Abstract
Federated learning (FL) lets hospitals and medical devices train a shared model without pooling patient records, but Internet of Medical Things (IoMT) deployments add three problems at once: devices run on small batteries, some participants may poison the model, and plain FL keeps no tamper-evident record of who contributed what. This paper presents BEA-FL, a blockchain-assisted, energy-aware and adaptive FL framework that handles these problems in one pipeline. Each round, clients are ranked by battery headroom, link quality, data volume and learned trust. Updates are weighted by a trust score that measures how well their direction and size agree with the coordinate-wise median update, and clearly inconsistent updates are rejected by validators. Only hashes and trust values are written to a hash-chained, quorum-signed ledger. We evaluated BEA-FL over 20 random seeds on three public clinical datasets against eight baselines, including FedProx, Median, Trimmed mean and Multi-Krum, under label-flipping, sign-flipping and noise-injection attacks. With 20% malicious clients on the Breast Cancer Wisconsin data, BEA-FL kept 96.9% accuracy where FedAvg fell to 90.1% (worst seed 73.5%), and it was within about one to two points of the strongest robust aggregators. It also used 5.5% less total energy on average and left 10.2 instead of 8.0 devices able to train after 30 rounds. The ledger added about 4 μs per record and flagged all 7,000 simulated tampering attempts. Protection weakened once 30% or more of the clients were malicious, a limit we report together with the other results.
Recommended Citation
Al-Helali, Adnan Hadi Mahdi
(2026)
"A Blockchain-Assisted Energy-Aware Adaptive Federated Learning Framework for Secure Internet of Medical Things Healthcare Systems,"
Mesopotamian Journal of Artificial Intelligence in Healthcare: Vol. 4:
Iss.
1, Article 9.
