In Federated Learning (FL), a central server coordinates a network of decentralized clients to train a shared machine learning model. The standard approach, Federated Averaging (FedAvg), operates in a simple loop: clients download the global model, perform several steps of local Stochastic Gradient Descent (SGD) on their private data, and upload their updated parameters to the server for averaging.
While computationally efficient, this paradigm encounters a fundamental mathematical obstacle when deployed in the real world: client drift.
The Geometry of Client Drift
Client drift occurs when clients possess heterogeneous, non-independently and identically distributed (non-IID) data. As noted by Jiang in a comprehensive survey on synchronous federated training, performing multiple local optimization steps causes individual client models to diverge significantly from the global model's optimal path.
To build intuition, imagine two clients training a image classifier. Client A's local dataset contains only images of cats, while Client B's dataset contains only dogs. Starting from a shared global model, Client A's local updates rapidly pull the weights toward a "cat-expert" local minimum. Simultaneously, Client B's updates pull the weights toward a "dog-expert" minimum.
When the server averages these drifted parameters, the resulting model does not land on a globally optimal solution that understands both animals. Instead, it suffers from a persistent optimization bias. As proved by Sun et al. through the lens of algorithmic stability, this data heterogeneity directly drives generalization error and destabilizes the entire training trajectory.
Control Variates as Gradient Anchors
To prevent local models from wandering off toward biased local minima, researchers introduce control variates to dynamically adjust local gradient steps.
In the foundational SCAFFOLD framework, Karimireddy et al. introduced a state variable that maintains a running estimate of the discrepancy between the global gradient direction and each client's local gradient direction. During local training, this control variate is added as a corrective offset to every single local step.
The control variate acts like an elastic tether. If a client's local data pushes the gradient strongly in a direction that conflicts with the global consensus, the control variate applies a counter-force. This continuous adjustment ensures that the local optimization trajectory remains aligned with the aggregate global objective, effectively neutralizing the impact of data heterogeneity.
This mechanism has been adapted across various domains to stabilize decentralized training:
- Algorithmic Extensions: Mangold et al. successfully applied control variates to linear stochastic approximation and reinforcement learning via SCAFFLSA.
- Communication Efficiency: To avoid the downlink overhead of broadcasting explicit control variables, Varno et al. developed AdaBest, which uses the implicit accumulation of standard model updates at the server to maintain an adaptive bias estimate.
- Asynchronous Systems: Zang demonstrated in FedAC that applying gradient correction continuously at every step of local SGD—rather than only at the end of a round—is vital for preventing asynchronous updates from drifting.
Server-Assisted Reference Gradients
An alternative strategy for anchoring local updates relies on a small, representative auxiliary dataset hosted directly on the central server.
Instead of relying entirely on clients to estimate the global direction, algorithms like SAFARI, proposed by Yang et al., use server-side data to compute a global reference gradient. This reference gradient acts as a corrective mechanism to steer the global model toward the true population mean. This is particularly valuable in scenarios with incomplete client participation, where hardware or connectivity constraints systematically exclude certain demographics of data from contributing updates.
Similarly, the FedCLG framework by Bian et al. leverages a server-side dataset to compute a reference gradient that counteracts the variance introduced by partial client participation. The server's gradient serves as a global compass, correcting the trajectory of the aggregated model even when the participating subset of clients is highly skewed.
Dual-Variable and Adaptive Corrections
As federated learning expands to more complex objectives, such as min-max optimization (common in adversarial training and Area Under the ROC Curve maximization), standard primal weight corrections become insufficient.
To address this, Guo et al. developed CODASCA, which introduces a dual-correction strategy. In min-max problems, client drift occurs simultaneously in both the primal model weights and the dual auxiliary variables. By maintaining control variates for both sets of variables, the authors ensure that the local saddle-point optimization remains globally aligned, achieving stable convergence without increasing communication frequency.
Furthermore, applying a uniform correction factor to all clients can lead to "over-correction" if a particular client's local update is already well-aligned with the global goal. To resolve this, Liu et al. introduced TACO, an adaptive correction framework. TACO personalizes the intensity of the gradient correction based on the directional similarity (dot product) between a client's local update and the global gradient. By dynamically scaling the global guidance, it prevents over-correction and stabilizes convergence across highly diverse participants.