A Secure Priority‐Adaptive Federated Learning Framework for Multimodal IoT Data Fusion in Real‐Time Edge Environments

Federated learning (FL)‐based models for real‐time IoT applications, in general, suffer from rigid architectures. They have limited adaptability to heterogeneous data and inadequate system‐level validation on real‐world hardware. Majority of the current implementations focus on structured sensing modalities such as physiological and environmental data. Our proposed framework is inherently modality‐invariant and can be extended to multimedia‐IoT scenarios involving audio, image, and video streams. The adaptive fusion and priority‐aware learning mechanisms are particularly suitable for handling heterogeneous multimodal data with varying reliability, bandwidth constraints, and latency sensitivity. In this paper, a generalized, priority‐adaptive FL framework is proposed and designed for robust, privacy‐preserving, and resource‐efficient analytics across broader IoT scenarios by combining dynamic fusion weights, differential privacy, and trust‐aware modules to support heterogeneous hardware, including NodeMCU‐based data source nodes and Raspberry Pi edge devices. The model is evaluated over different datasets such as MIMIC‐III and UCI Air Quality datasets. Results show that the proposed framework achieves up to 99.72% accuracy and an F1‐score of 99.73%. The energy consumption is maintained below 8.1 mJ throughout the evaluation rounds, and inference latency is approximately 78 ms. To illustrate the practical feasibility of the proposed model, a hardware prototype for air quality monitoring is designed and validated across various system metrics, such as the Privacy Deviation Index (0.02) under real‐world conditions. These results validate the robustness and scalability of the method and assert it as a very effective solution for a wide variety of critical IoT and Edge learning‐related applications.