Privacy-Preserving Federated Unlearning with Ontology-Guided Relevance Modeling for Secure Distributed Systems

Federated Learning (FL) is a privacy-focused technique for training models; however, most existing unlearning techniques in FL fall significantly short of the efficiency and situational awareness required by the GDPR. The paper introduces two new unlearning methods: EG-FedUnlearn, a gradient-based technique that eliminates the effect of specific target clients without retraining, and OFU-Ontology, an ontology-based approach that ranks data importance to facilitate forgetting contextually. EG-FedUnlearn directly eliminates the contributions of specific target data by reversing the gradient, whereas OFU-Ontology utilizes semantic relevance to prioritize forgetting data of the least importance, thereby minimizing the unlearning-induced degradation of models. The results of experiments on seven benchmark datasets demonstrate the good performance of both algorithms. OFU-Ontology yields 98% accuracy of unlearning while maintaining high model utility with very limited accuracy loss under class-based deletion on MNIST (e.g., 95%), surpassing FedEraser and VeriFi on the multiple metrics of residual influence, communication overhead, and computational cost. These results indicate that the cooperation of efficient unlearning algorithms with semantic reasoning, minimized unlearning costs, and operational performance in a distributed environment. This paper becomes the first to incorporate ontological knowledge into federated unlearning, thereby opening new avenues for scalable and intelligent private machine learning systems.