Federated Learning Model Privacy and Accuracy Trade-off

The patent divides the training data into multiple clients, each holding data from a single individual. This segmentation enables federated learning where each client trains locally on its own data without sharing personal information, thus protecting privacy while contributing to the global model. The model is segmented into local individual models and a global common model that aggregates knowledge from all clients.

The patent introduces a server as an intermediary that coordinates the federated learning process. The server aggregates local models from individual clients to create an updated common model, which is then distributed back to clients. This intermediary structure enables privacy protection by preventing direct data sharing while maintaining authentication accuracy through coordinated model aggregation and updates.

The patent creates a common model that serves multiple functions: it acts as an individual model for each client while simultaneously serving as a universal model for all clients. This common model is trained on data from multiple individuals through federated learning, enabling it to perform class comparison and authentication across different individuals without requiring direct access to their personal data.

The patent merges local individual models from multiple clients into a single common model through the federated learning process. Each client contributes its local knowledge to the global model, which combines these diverse data sources while maintaining privacy. The merged common model achieves accurate classification by learning from the combined distribution of data across all clients.

The patent inverts the traditional approach by having the server hold the common model instead of clients holding representative vectors. The server uses the common model to generate individual models for each client, which are then stored locally. This inversion improves security because the sensitive common model remains on the server, while clients only possess their individual models that are less informative for impersonation attacks.

The patent extracts the sensitive representative vectors from the clients and consolidates them into a common model on the server. This extraction removes the security vulnerability where clients would otherwise store and potentially leak their individual representative vectors. The server securely manages the aggregated model, reducing the attack surface for impersonation while maintaining authentication accuracy.