Continual Federated Learning

The primary failure mode that continual learning aims to solve. When a neural network is trained sequentially on new data distributions, it exhibits a rapid, often irreversible degradation in performance on previously learned tasks. In a federated context, this occurs as the global model adapts to a stream of non-stationary client data, overwriting old knowledge. Mitigation strategies include elastic weight consolidation (EWC), which penalizes changes to parameters critical for past tasks, and experience replay, which interleaves synthetic or buffered historical data with new updates.

The divergence of local model updates from the hypothetical global optimum due to non-IID data distributions across clients. In continual settings, drift is exacerbated because client data distributions also shift over time, creating a moving target for the global model. This statistical heterogeneity leads to unstable convergence and can accelerate catastrophic forgetting. FedProx and SCAFFOLD are optimization frameworks that introduce correction terms to combat client drift without compromising the privacy guarantees of the federated averaging process.

The foundational optimization algorithm that underpins most federated learning systems. FedAvg works by performing multiple local stochastic gradient descent (SGD) steps on client devices and then averaging the resulting model weights on a central server. While communication-efficient, standard FedAvg assumes static data distributions and struggles with the temporal concept drift inherent in continual learning. Extensions like FedProx and FedCurv adapt the averaging paradigm to sequential, non-stationary environments.

A data distribution scenario where local client datasets are not independently and identically distributed. This is the norm in real-world federated deployments, not the exception. In continual learning, non-IIDness has a temporal dimension: the label distribution, feature space, or concept definition on a single client can shift over time. Handling this dual heterogeneity—spatial across clients and temporal within a client stream—is the central challenge of continual federated learning.

A model compression technique where a smaller 'student' model is trained to mimic the output distribution of a larger 'teacher' model. In continual federated learning, distillation serves a dual purpose:

• Global-to-local: Distilling a global model into a personalized on-device model without sharing raw data.
• Anti-forgetting: Using the previous global model as a teacher to regularize the current training round, preventing the student from diverging too far from historically learned representations. This provides a privacy-compatible alternative to storing raw exemplars for replay.

A cryptographic protocol that allows a central server to compute the sum of model updates from multiple clients without being able to inspect any individual client's contribution. This is critical for continual federated learning because the sequential nature of updates over time creates more opportunities for gradient leakage attacks to reconstruct private training data. Secure aggregation ensures that even the temporal trajectory of a single client's model drift remains cryptographically obscured from the coordinating server.