Federated continual learning: A comprehensive survey on lifelong and privacy-preserving learning over distributed and non-stationary data
A comprehensive survey examines Federated Continual Learning (FCL), which combines federated learning's privacy-preserving distributed training with continual learning's ability to adapt to evolving data. The research addresses a critical gap in current FL systems that assume static data, proposing frameworks for real-world applications like healthcare and IoT where data streams continuously shift, causing performance degradation and catastrophic forgetting.
Federated Learning has revolutionized collaborative model training by enabling multiple distributed clients to train shared models without exposing raw data. However, existing FL implementations operate under a fundamental assumption: data remains stationary across training rounds. This assumption collapses in real-world deployments where data distributions shift continuously over time—a phenomenon known as temporal drift or concept drift. The survey identifies this disconnect as a critical limitation affecting healthcare systems, industrial IoT networks, cybersecurity platforms, and smart city infrastructure.
Continual Learning provides theoretical and practical solutions for handling non-stationary data streams through techniques like replay buffers, regularization strategies, and dynamic architectures. Yet classical CL research has largely ignored the federated context's unique constraints: privacy requirements, communication bandwidth limitations, and client heterogeneity. Federated Continual Learning bridges this gap by developing methods that maintain privacy guarantees while supporting lifelong learning across distributed, heterogeneous clients experiencing different data distributions.
The survey's taxonomy and systematic analysis directly impact ML practitioners deploying systems in evolving environments. Organizations building intelligent infrastructure must now reconcile three competing demands: privacy protection, communication efficiency, and adaptation to drift. The identification of open challenges—extreme heterogeneity under temporal drift, scalable privacy-preserving memory mechanisms, and standardized benchmarks—reveals where innovation is most urgent.
Looking forward, standardized FCL benchmarks will enable reproducible research and accelerate adoption. Solutions addressing extreme heterogeneity and communication-efficient continual learning mechanisms represent the next frontier. Success here determines whether federated systems can reliably support long-term autonomous operations in dynamic real-world environments.
- →Federated Learning systems fail under non-stationary data conditions found in real-world healthcare, IoT, and cybersecurity applications due to catastrophic forgetting and performance degradation.
- →Federated Continual Learning combines privacy-preserving distributed training with adaptive learning mechanisms designed for evolving data distributions across heterogeneous clients.
- →Current research lacks standardized benchmarks and evaluation metrics for assessing long-term FCL performance, hindering reproducible research and practical deployment.
- →Key open challenges include managing extreme client heterogeneity during temporal drift, designing scalable privacy-aware memory systems, and establishing industry benchmarks.
- →FCL frameworks are essential for enabling truly deployable autonomous systems in dynamic environments like smart cities, industrial automation, and healthcare monitoring.