RIFT: A Secure and Adaptive Reinforced Federated Architecture for AI‐Driven Industrial Medical Sensor Networks

ABSTRACT

AI‐driven medical sensor networks require sensing tasks, distributed learning, and communication resources to be coordinated across heterogeneous and potentially unreliable devices. We present RIFT, a round‐based architecture that links medical crowdsensing with federated learning through quality‐aware aggregation within a unified workflow. RIFT uses a medical sensing market for feasible task–agent matching and formulates cross‐round resource and learning control as a constrained Markov decision process. The architecture further integrates truth and quality estimation with anomaly‐aware federated aggregation, Top‐ update sparsification, and lightweight audit records. In the aggregation process, sensing‐quality and reliability information is combined with update‐level anomaly measures to limit the influence of unreliable participants. We implement the federated learning, quality‐gating, communication‐compression, and audit paths using Flower and PyTorch and evaluate them on CIFAR‐10 and NEU‐CLS as controlled proxy benchmarks. Experiments consider non‐IID partitions, composite poisoning attacks, optimiser changes, and different communication sparsification ratios. In the system‐level comparison, RIFT maintains higher accuracy than FedSense under the evaluated poisoning configurations, with particularly large performance differences on NEU‐CLS. The compression experiments further characterise the trade‐off between communication reduction and predictive performance. These results demonstrate the robustness and communication behaviour of the implemented RIFT components.