Federated Out-of-Distribution Detection (OOD Detection)

Each institution computes a local OOD score for every inference sample using methods like Mahalanobis distance or energy-based models on its own data partition. These scores are then securely aggregated across the network using federated averaging or secure aggregation protocols to produce a global anomaly threshold, ensuring no raw patient features leave the local site during the safety check.

Federated OOD detection must distinguish between multiple types of distributional shift:

• Covariate Shift: The input feature distribution changes (e.g., a new MRI scanner vendor) while the diagnostic task remains the same.
• Semantic Shift: Entirely new classes appear at inference that were absent from all federated training nodes (e.g., a novel pathology).
• Label Shift: The prevalence of known conditions changes across participating hospitals, requiring recalibration of OOD thresholds.

A powerful OOD signal in federated systems is the disagreement between local and global model predictions. When a sample is in-distribution for the global model but anomalous for a specific local model, it may indicate subpopulation shift or domain-specific outliers. This discrepancy is computed locally and aggregated via federated uncertainty quantification methods to avoid centralizing prediction vectors.

Traditional open set recognition requires access to known class prototypes. In federated settings, class-conditional feature centroids are computed locally and shared through differentially private mechanisms with a calibrated privacy budget (epsilon). The global OOD detector then measures the distance between a new sample's embedding and these protected centroids, flagging inputs that fall outside a statistically defined radius as out-of-distribution.

Energy-based OOD detection uses a scalar energy score—lower for in-distribution, higher for OOD—derived from the model's logits. In federated settings, each client trains a local energy function on its data partition. The global energy threshold is aggregated using FedAvg on the energy parameters, creating a consensus boundary that respects the heterogeneity of clinical data distributions across institutions.

When a federated OOD detector flags an input, configurable rejection policies determine the clinical workflow:

• Hard Rejection: The model abstains entirely, routing the case for mandatory human review.
• Soft Rejection with Confidence: The prediction is delivered alongside a calibrated Expected Calibration Error (ECE) warning and an OOD flag.
• Deferral to Local Expert: The sample is redirected to a site-specific model trained on that institution's unique subpopulation, leveraging personalized federated learning.

Methods such as Monte Carlo Dropout or Deep Ensembles applied in a federated setting to estimate the epistemic uncertainty of a model's predictions. This is the foundational signal for OOD detection, as inputs far from the training distribution typically yield high model uncertainty. In a federated context, uncertainty must be computed locally and aggregated without sharing raw confidence scores, often using secure aggregation protocols.

A change in the statistical relationship between input features and the target variable over time. In federated healthcare deployments, concept drift can manifest when a new imaging device produces subtly different scans or when clinical protocols evolve at one institution. Federated OOD detection systems must distinguish between benign covariate shift and true concept drift to avoid flagging valid clinical variations as out-of-distribution.

The continuous monitoring of a deployed federated model's performance to identify degradation caused by concept drift or data drift in distributed input streams. Key techniques include:

• Federated Population Stability Index (PSI) for feature distribution monitoring
• Federated Expected Calibration Error (ECE) for confidence reliability
• Decentralized statistical process control charts These methods operate without centralizing live inference data, preserving patient privacy during ongoing surveillance.

A distribution-free framework for generating prediction intervals with a guaranteed coverage probability. Unlike traditional OOD detection, conformal prediction provides rigorous statistical guarantees without assuming a specific data distribution. In federated settings, calibration scores are computed locally and aggregated to produce valid prediction sets that automatically flag inputs where the model cannot make a reliable prediction.

The capability of a federated model to perform accurately on entirely new, unseen client domains that were not present during training. This is the proactive counterpart to OOD detection—rather than merely flagging unfamiliar inputs, domain generalization techniques such as invariant risk minimization and federated adversarial training aim to build models that are inherently robust to distributional shift across diverse healthcare systems.

A privacy audit technique that simulates an adversary attempting to determine if a specific patient's record was used in a federated training run. This is closely related to OOD detection because both rely on analyzing model confidence signals—an OOD detector's confidence thresholds can inadvertently leak membership information if not carefully calibrated with differential privacy guarantees.