Continual Federated Learning for Dynamic Data Environments

Federated learning lets smartphones, computers, or other such end devices learn without transmitting their raw data. Instead of sending personal data to a common server, each device does its own learning locally. Consequently, updated parameters from all devices are transferred to a central server to improve a global model for classification purposes. This approach ensures the privacy of individual data by keeping it decentralized (on device) and still allowing the models to learn collectively. Federated learning models often face the challenge of becoming outdated once training is complete, particularly when confronted with new data as the model is static. To avoid this issue, we incorporate continual learning into federated learning, which allows the devices' data to be sent as streams. This ensures that the global model remains adaptable, continually incorporating novel information without forgetting what it has learned in the past. The results of the study suggest that such a combined approach might be more effective than conventional federated learning methods. Twelve experiments were carried out applying continual learning to a federated learning setup for image classification tasks using the 10-fold cross-validation method on the MNIST, Fashion-MNIST and CIFAR-10 datasets; thus, the performance of the framework is calculated.