On the Convergence of Continual Federated Learning Using Incrementally Aggregated Gradients

Satish Kumar Keshri, Nazreen Shah, Ranjitha PrasadAISTATS 2025

Introduction

In the evolving landscape of distributed machine learning, Continual Federated Learning (CFL) has emerged as a critical paradigm. It addresses the challenge of learning from non-stationary data streams across decentralized edge devices without compromising data privacy. However, CFL is plagued by catastrophic forgetting, where models lose performance on previously learned tasks as they adapt to new ones. This paper introduces C-FLAG (Continual Federated Learning with Aggregated Gradients), a framework designed to balance the stability-plasticity trade-off while providing rigorous theoretical guarantees.

The C-FLAG Framework

C-FLAG integrates several innovative components to mitigate forgetting and ensure efficient convergence in distributed environments.

Memory-Based Replay and Incremental Gradients

To retain knowledge of past tasks, each client maintains a small episodic memory buffer. Unlike traditional replay methods that require extensive retraining, C-FLAG performs a single gradient step on this memory for every $E$ steps on the current task. To further reduce computational overhead, the framework utilizes Incremental Aggregated Gradients (IAG), leveraging delayed gradients from previous iterations to provide a variance-reduction effect without the cost of full gradient computation at every step.

Adaptive Learning Rates (AdapLR)

The core of C-FLAG's performance lies in its Adaptive Learning Rate (AdapLR) strategy. By formulating a sub-optimization problem, the algorithm distinguishes between client updates that cause Interference (harming old tasks) and those that facilitate Transference (benefiting old tasks). AdapLR dynamically adjusts local learning rates to maximize positive transference and suppress interference, effectively guiding the global model toward regions of low loss for both current and historical tasks.

Figure 3: Overview of the C-FLAG system architecture and the optimization landscape illustrating the aggregated update strategy.

Theoretical Convergence

A significant contribution of this work is the first $O(1/\sqrt{T})$ convergence guarantee for non-convex CFL. The authors prove that by employing diminishing learning rates for current tasks while maintaining constant rates for memory updates, the cumulative forgetting can be bounded, ensuring the model converges to a stationary point even under complex, non-stationary data distributions.

Experimental Results

The researchers evaluated C-FLAG against several baselines, including EWC-FL and Erg-FL, using standard benchmarks: Split-CIFAR10, Split-CIFAR100, and Split-TinyImageNet.

Performance and Retention

C-FLAG consistently outperformed existing methods in both accuracy and forgetting metrics. For instance, on the Split-CIFAR100 benchmark, C-FLAG achieved an average accuracy of 66.85%, a substantial improvement over EWC-FL's 50.48%. More importantly, it maintained a remarkably low forgetting rate of 7.30%, whereas competing methods like Erg-FL suffered from forgetting rates as high as 43.66%.

Figure 5: Accuracy comparison across sequential tasks on Split-CIFAR10 (left) and Split-CIFAR100 (right), demonstrating C-FLAG's superior performance retention.

Robustness to Heterogeneity

The framework's effectiveness was further validated across varying levels of data heterogeneity. C-FLAG demonstrated superior resilience to non-IID data distributions, a common hurdle in real-world federated learning deployments. Ablation studies also confirmed that the model scales effectively with an increasing number of clients and benefits from larger memory buffers when handling complex datasets.

Conclusion

C-FLAG represents a significant step forward in Continual Federated Learning, offering a robust solution to catastrophic forgetting with strong theoretical backing. By intelligently managing gradient updates and local learning rates, it enables edge devices to learn continuously and collaboratively. Future work will explore extending these techniques to asynchronous federated settings and extremely non-stationary environments.

Summary updated

BibTeX
@inproceedings{keshri2025convergence,
  author = {Satish Kumar Keshri and Nazreen Shah and Ranjitha Prasad},
  title = {{On the Convergence of Continual Federated Learning Using Incrementally Aggregated Gradients}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2025},
}
Figure