Federated Drift Detection

Before aggregation, the central server applies statistical tests to the distribution of local model updates to detect anomalous clients. Techniques like the Kolmogorov-Smirnov test or Maximum Mean Discrepancy (MMD) compare the current round of weight updates against a historical reference distribution. A significant p-value indicates that a client's underlying data distribution has shifted, flagging it for exclusion or triggering a selective retraining protocol. This method detects drift without inspecting raw data.

Each client continuously evaluates a held-out local validation set and reports key metrics—accuracy, F1-score, log-loss—to the aggregation server. Drift is declared when a client's performance trajectory deviates from the global median by a predefined threshold. This approach directly captures concept drift where the relationship between features and labels changes. For example, if a hospital's diagnostic model accuracy drops by 5% while others remain stable, it signals a localized shift in patient presentation or labeling practices.

Clients compute compact statistical summaries of their data representations—such as the mean and covariance of the penultimate layer's activations—and share these with the server. The server monitors the Fréchet distance or Wasserstein distance between these embeddings over time. A sudden increase in distance indicates covariate shift or feature distribution skew. This technique is particularly effective for detecting silent data corruption or changes in medical imaging acquisition protocols without transmitting images.

Drift detection is only valuable when coupled with automated remediation. When drift exceeds a critical threshold, the system can trigger one of several responses:

• Selective Fine-Tuning: Only the drifted client performs additional local epochs.
• Cluster Reassignment: The client is moved to a new cluster with similar data distributions.
• Global Model Rollback: The server reverts to a previous stable checkpoint.
• Human-in-the-Loop Alert: A notification is sent to a clinical ML engineer for manual investigation. These triggers prevent a single drifted client from degrading the global model.

Each client computes a local data quality score based on metrics like missing value ratios, feature variance, and label consistency. These scores are aggregated to establish a network-wide quality baseline. A client whose quality score drops below the baseline by more than two standard deviations is flagged for drift. This method detects data poisoning and sensor degradation in medical IoT devices, ensuring that faulty edge hardware does not corrupt the collaborative training process.

The aggregation server maintains a time-series database of global model parameters and client contributions. It applies change-point detection algorithms like CUSUM or Bayesian Online Change-Point Detection to identify abrupt shifts in the parameter space. This approach distinguishes between gradual, benign distribution evolution and sudden, harmful drift. For instance, a sudden shift in a radiology model's convolution filters might indicate a new scanner vendor being introduced at a site, requiring federated domain adaptation.

The underlying phenomenon where the statistical relationship between input features and target labels changes over time. In a clinical context, this occurs when diagnostic criteria evolve, new disease variants emerge, or treatment protocols alter patient outcomes. Federated drift detection systems must distinguish between genuine concept drift and the noise introduced by non-IID data distributions across hospitals.

• Real drift: Changes in P(Y|X), such as a new biomarker redefining a disease
• Virtual drift: Changes in P(X) without affecting the decision boundary
• Recurring contexts: Seasonal patterns like flu outbreaks that temporarily shift distributions

The natural variation in data distributions across federated clients that complicates drift detection. A shift detected in the global model may originate from a single hospital changing its patient intake demographics rather than a true population-wide drift. Federated drift detection must decompose global signal changes into local and systemic components.

• Feature distribution skew: Different hospitals use different MRI machine vendors
• Label distribution skew: A specialist cancer center sees disproportionately more malignant cases
• Temporal heterogeneity: Clients join and leave the federation at different times

The umbrella term for the mismatch between training and deployment distributions in a federated network. Unlike centralized drift detection, federated dataset shift requires monitoring without direct access to raw patient data. Techniques rely on comparing local model update statistics, gradient norms, or prediction entropy across rounds.

• Prior probability shift: Changes in class prevalence at individual sites
• Covariate shift: New patient populations with different demographic profiles
• Open-set shift: Novel disease classes appearing that were absent during training

The architectural paradigm that enables models to adapt to detected drift without catastrophic forgetting. When federated drift detection triggers a retraining signal, the system must incorporate new data distributions while preserving performance on previously learned clinical tasks. This is critical for longitudinal patient monitoring systems.

• Elastic weight consolidation: Protecting parameters important for prior tasks
• Memory replay: Using synthetic or proxy data to rehearse old distributions
• Dynamic architecture expansion: Adding capacity for new concepts without disrupting existing knowledge

The auditing framework required to validate that drift detection mechanisms are functioning correctly across all nodes. Without centralized test data, federated drift detection systems must rely on distributed evaluation protocols that compute performance metrics locally and share only aggregated statistics.

• Federated AUC monitoring: Tracking receiver operating characteristic shifts per site
• Prediction entropy tracking: Rising uncertainty often precedes performance degradation
• Local loss divergence: Comparing individual client loss trajectories to the federation median

The process of quantifying each client's contribution to model stability and drift resilience. When federated drift detection identifies a degrading client, data valuation helps determine whether to down-weight, retrain, or isolate that node. Game-theoretic approaches like the Shapley value can identify which hospital's data is causing distributional instability.

• Marginal contribution analysis: Measuring how each client affects global model loss
• Leave-one-out retraining: Assessing impact by excluding suspect clients
• Data quality scoring: Assigning drift-risk scores to individual institutional datasets