Trust Based Federated Learning for Privacy-preserving in Connected Smart Communities
S Difrina, Mahalingam Ramkumar, G. S. R. Emil Selvan · 2025
Federated Learning (FL) has become a prominent privacy-preserving method for training machine learning models across distributed edge devices. The FL frameworks frequently function under the uniform trust assumption, which assumes that all participating nodes have the same level of security and dependability. This presumption ignores how heterogeneous IoT ecosystems are by nature, with different computing power, security and reliability impacting the training models. Significant vulnerabilities arise by the presence of resourceconstrained devices with varying performance characteristics and a variety of risk profiles. A trust based federated learning framework is proposed for audio classification that combines data partitioning based on Jaccard similarity and dynamic trustaware node selection. In each training round, the trust score is calculated based on the different parameters, including data breaches, reputation decay, security compliance, communication efficiency, contribution correctness and confidence factor. To guarantee a safe and efficient aggregation process, only nodes with trust scores greater than 0.7 are allowed to take part in the federated learning. Furthermore, the effective data splitting technique based on Jaccard similarity guarantees representative and varied data partitions for efficient model training. For audio categorisation, a modified ResNet34 has been used to adjust the inputs from 1 -channel spectrograms. Compared to the conventional FL approaches, experimental results shows that the proposed trust aware federated learning improves training speed, increases model robustness and reduces the impact of malicious or unreliable nodes. The proposed work highlights the potential of trust-based FL in privacy-sensitive IoT applications, paving the way for more secure and efficient decentralized learning frameworks.