Client Drift

The most direct measure of client drift is the increasing Euclidean distance between a client's local model weights and the aggregated global model weights after each round of local training. In non-IID settings, local SGD steps move parameters toward local minima that minimize empirical risk on the client's own skewed data distribution, rather than the true global loss landscape. This creates a centrifugal effect where each client's update vector points in a direction that conflicts with other clients. The divergence can be quantified using layer-wise cosine similarity or L2 norm differences, with deeper layers often exhibiting greater drift due to their role in learning dataset-specific representations.

As local training progresses across communication rounds, the variance of stochastic gradients across clients increases substantially. This gradient dissimilarity is a root cause of client drift and can be measured by the gradient diversity metric—the average pairwise cosine dissimilarity between client update vectors. High gradient variance indicates that clients are optimizing toward incompatible local objectives. This phenomenon is exacerbated by:

• Label distribution skew: Some clients have disproportionate representation of certain classes
• Feature distribution skew: Clients' input features follow different statistical distributions
• Quantity skew: Significant variance in local dataset sizes

The result is a noisy aggregated update that oscillates rather than converging smoothly.

A subtle but critical characteristic of client drift is the diminishing magnitude of local update vectors from clients with highly skewed data distributions. These clients' models converge quickly to their local optima, producing updates with small L2 norms that contribute minimally to the global aggregation. This creates a silent majority problem where the global model is disproportionately influenced by a few clients with large, noisy updates while drifting clients become effectively excluded from the collaborative learning process. Monitoring the coefficient of variation of update norms across clients serves as an early warning indicator of this imbalance.

Client drift manifests structurally as dimensional collapse in the learned feature representations. When clients train on narrow, non-overlapping data distributions, their models learn compressed, low-rank representations that fail to generalize. This can be detected by analyzing the singular value spectrum of the penultimate layer's weight matrix—a steep drop in effective rank indicates that the model has discarded feature dimensions relevant to other clients' data. Centered Kernel Alignment (CKA) between clients' representation layers provides a pairwise similarity measure, with low CKA scores revealing that clients are learning fundamentally different feature extractors rather than complementary ones.

The practical consequence of client drift is uneven model performance across the federated population. While the global aggregated model may show acceptable average accuracy, per-client evaluation reveals a bimodal distribution: clients with data distributions close to the global mean perform well, while outlier clients suffer significant accuracy degradation. This is measured by the standard deviation of per-client test accuracy and the worst-client accuracy gap. In healthcare federated learning, this translates to a model that works well for typical patient populations but fails for underrepresented demographic groups or rare clinical presentations—directly violating fairness requirements.

Client drift directly increases the number of communication rounds required for convergence, measured by the round-to-accuracy metric. In severely non-IID settings, the global model may require 3-10x more communication rounds to reach the same target accuracy compared to an IID baseline. This inefficiency arises because conflicting client updates partially cancel each other out during aggregation, forcing the optimization process to take smaller effective steps toward the global optimum. The drift penalty factor—the ratio of rounds needed in non-IID versus IID settings—quantifies this overhead and is a key metric for evaluating drift mitigation strategies like FedProx, SCAFFOLD, or FedDyn.

The continuous monitoring of a deployed federated model's performance to identify degradation over time caused by concept drift or data drift in distributed input streams. Unlike centralized monitoring, this process must compute statistical divergence metrics—such as the Federated Population Stability Index (PSI)—without aggregating raw inference data. When drift is detected, it triggers automated retraining or model rollback protocols to maintain clinical safety.

A change in the statistical relationship between input features and the target variable over time. In a federated healthcare network, this might occur when a new treatment protocol alters the correlation between lab values and patient outcomes at one hospital. Unlike data drift (a shift in input distributions), concept drift directly invalidates the model's learned decision boundary, requiring detection via Federated Uncertainty Quantification or degradation in Expected Calibration Error (ECE).

The process of auditing a federated model for unfair performance disparities across protected groups by securely computing fairness metrics on distributed validation data. Key metrics include:

• Demographic Parity: Equal positive prediction rates across groups
• Equalized Odds: Equal true positive and false positive rates across groups These computations must be performed without exposing individual patient demographics, often using Secure Aggregation to sum confusion matrix components.

A set of techniques designed to interpret model predictions in a decentralized setting, providing auditable feature attribution without exposing local patient data. Federated Shapley Values decompose a prediction into per-feature contributions by securely aggregating marginal contribution calculations across institutions. Federated LIME approximates local decision boundaries using interpretable surrogate models trained on perturbed samples, all while keeping raw records at their origin.

A privacy-preserving method for aggregating classification performance metrics across multiple institutions by securely summing true positive, false positive, true negative, and false negative counts. Each client computes its local confusion matrix on its private test set, then shares only these four aggregate integers via Secure Aggregation (SecAgg). The global matrix enables computation of Federated F1-Score, Federated AUC, sensitivity, and specificity without exposing individual predictions.

A privacy audit technique that simulates an adversary attempting to determine if a specific patient's record was used in a federated training run. By analyzing the model's prediction confidence on target records versus non-training records, auditors can quantify information leakage. This evaluation is critical for validating the effectiveness of Differential Privacy guarantees and ensuring that the Privacy Budget (Epsilon Budget) has not been exhausted.