Abstract
Federated learning (FL) often struggles when clients possess highly diverse data distributions, a phenomenon known as "client drift" that hampers global model convergence. While existing variance reduction methods like SCAFFOLD mitigate this drift, they typically double the communication overhead, making them impractical for bandwidth-constrained environments. This paper introduces FedPVR (Partial Variance Reduction), a method that applies gradient correction selectively to the final classification layers of deep neural networks. We demonstrate that early feature extraction layers are naturally robust to heterogeneity, and that stabilizing only the "head" of the model achieves up to a 6.7x speedup over standard FedAvg with significantly lower communication costs.
Introduction
Federated Learning (FL) enables collaborative training of machine learning models across decentralized clients without raw data exchange. The standard algorithm, FedAvg, performs well under independent and identically distributed (IID) data. However, in real-world scenarios where data is heterogeneous (non-IID), local updates diverge from the global objective. This "client drift" slows convergence and degrades final accuracy.
Previous attempts to solve this using full variance reduction (VR) require transmitting control variates for every model parameter. For deep networks, this doubling of communication volume is often prohibitive. We propose a more efficient alternative based on the observation that drift is not uniform across model layers.
The Empirical Basis for Partial Correction
To understand where drift occurs, we analyzed deep models using Centered Kernel Alignment (CKA) and a novel "Drift Diversity" metric. Our findings reveal a stark contrast between early and late layers: feature extraction layers maintain high similarity across clients even under extreme heterogeneity, whereas the final classification layers exhibit significant divergence.
This insight suggests that correcting the entire model is unnecessary and potentially counterproductive. By allowing early layers to remain "diverse," they can learn richer representations that benefit the global model, provided the final decision-making "head" is stabilized.
FedPVR: Partial Variance Reduction
FedPVR applies variance reduction selectively. It maintains control variates—states that track the difference between local and global update directions—only for a subset of the model's final layers.
By using a binary mask to restrict correction to the "head," FedPVR reduces the communication overhead from the 4x required by SCAFFOLD to approximately 2.1x relative to FedAvg, while maintaining or exceeding the convergence speed of full-model correction.
Experimental Evaluation
We evaluated FedPVR on CIFAR-10 and CIFAR-100 datasets using VGG-11 and ResNet-8 architectures. Data was partitioned using a Dirichlet distribution ($\alpha$) to simulate varying degrees of heterogeneity.
Key Results
- Performance Gains: FedPVR achieved a 1.5x to 6.7x speedup over FedAvg. On CIFAR-100 with extreme heterogeneity ($\alpha=0.1$), FedPVR reached 43.5% accuracy, outperforming FedAvg (34.3%) and FedProx (35.0%).
- Communication Efficiency: FedPVR achieves competitive accuracy with roughly half the extra data transmission required by state-of-the-art variance reduction methods.
- Regularization Effect: In moderate heterogeneity settings, FedPVR occasionally outperformed centralized training, suggesting that the stabilized distributed updates provide a beneficial regularizing effect.
Conclusion
FedPVR demonstrates that targeted variance reduction is a powerful and efficient strategy for heterogeneous federated learning. By focusing correction on the classification "head," we resolve client drift while preserving the beneficial feature diversity of early layers. Future work includes developing adaptive methods to automatically select the optimal layers for correction across varying architectures and data distributions.
Summary updated