Enhancing the Resilience of a Federated Learning Global Model Using Client Model Benchmark Validation

Federated learning (FL) makes it possible for users to share trained models with one another, thereby removing the necessity of publicly centralizing training data. One of the best and most cost-effective ways to connect users is through email. To steal sensitive information, spam emails might trick users into visiting malicious websites or performing other fraudulent actions. The developed semantic parser creates email metadata datasets from multiple email corpuses and populates the email domain ontology to facilitate the privacy of the information contained in email messages. There is a new idea to make FL global models more resistant to Byzantine attacks. It involves accepting updates only from strong participants whose local model shows higher validation scores using benchmark datasets. The proposed approach integrates FL, the email domain-specific ontology, the semantic parser, and a collection of benchmark datasets from heterogeneous email corpuses. By giving meaning to the metadata of an email message, the email’s domain-specific ontology made it possible to create datasets for email benchmark corpuses and participant updates in a unified format with the same features. In order to avoid fraudulently modified client updates from being applied to the global model, the experimental results approved the proposed approach to strengthen the resiliency of an FL global model by utilizing client model benchmark validation.