Federated Bias Detection

A group fairness metric requiring that a model's positive prediction rate is statistically equal across all protected demographic groups. In a federated context, this is computed by securely aggregating the count of positive predictions and total predictions for each group from every client node without sharing individual-level outcomes. Also known as statistical parity, this metric ensures that the probability of receiving a positive diagnostic recommendation is independent of group membership. A key limitation is that it does not account for legitimate differences in base rates between populations, which can lead to label flipping if applied naively.

A stricter fairness criterion requiring that a model has equal true positive rates (TPR) and false positive rates (FPR) across all protected groups. In federated evaluation, this demands the secure computation of a federated confusion matrix for each demographic slice, aggregating true positives, false positives, true negatives, and false negatives across institutions. Equalized odds ensures that the model is equally accurate for all groups in both positive and negative prediction classes, addressing the key shortcoming of demographic parity by conditioning on the true outcome.

A relaxed variant of equalized odds that requires only equal true positive rates across protected groups. This metric ensures that patients who should receive a positive diagnosis have an equal chance of receiving it regardless of their demographic group. In a federated setting, this is computed by securely summing true positives and total actual positives for each group across all participating hospitals. This metric is particularly relevant in healthcare where ensuring equitable access to treatment is the primary concern, while differences in false positive rates may be tolerated.

A calibration-based fairness metric requiring that the positive predictive value (PPV) — the probability that a positive prediction is correct — is equal across demographic groups. In federated evaluation, this requires aggregating true positives and total positive predictions per group across all nodes. Predictive parity ensures that when a model flags a patient for intervention, the likelihood that the flag is correct does not depend on group membership. This is critical in clinical settings where false alarms can lead to unnecessary and potentially harmful follow-up procedures.

The foundational privacy-preserving data structure for computing all group fairness metrics in a decentralized network. Each client computes a local confusion matrix stratified by protected attribute and transmits only the aggregate counts — true positives, false positives, true negatives, and false negatives — to the central server. These counts are summed across all institutions using secure aggregation to produce a global confusion matrix per demographic group. This approach enables full fairness auditing without exposing any individual patient's prediction or group membership.

A quantitative measure of demographic parity defined as the ratio of the positive prediction rate for the unprivileged group to that of the privileged group. A ratio of 1.0 indicates perfect parity, while values below 0.8 are commonly flagged as potential discrimination under the four-fifths rule used in U.S. employment law. In federated settings, this ratio is computed from securely aggregated group-level prediction counts. The metric provides a single, interpretable number for regulatory reporting but does not distinguish between justified and unjustified disparities.

A group fairness metric requiring that a model's positive prediction rate is equal across different demographic groups. In a federated context, this is computed by securely aggregating the count of positive predictions and total predictions for each group from all nodes without sharing individual outcomes. A model satisfies demographic parity if P(ŷ=1|A=a) = P(ŷ=1|A=b) for all protected groups A. While widely used, it does not account for ground truth labels, which can mask unfairness if base rates differ across groups.

A stricter fairness criterion than demographic parity, requiring that a model has equal true positive rates and false positive rates across protected groups. In federated evaluation, this demands secure computation of a federated confusion matrix per group. A model satisfies equalized odds if P(ŷ=1|Y=1, A=a) = P(ŷ=1|Y=1, A=b) AND P(ŷ=1|Y=0, A=a) = P(ŷ=1|Y=0, A=b). This metric ensures that errors are distributed equally, making it particularly relevant for clinical decision support where both missed diagnoses and false alarms carry significant consequences.

A privacy-preserving method for aggregating classification performance 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 validation data, then shares only the aggregate counts via secure aggregation or differential privacy mechanisms. The central server sums these counts to construct a global confusion matrix, enabling the computation of precision, recall, F1-score, and fairness metrics without exposing any individual patient's prediction or outcome.

A decentralized implementation of Shapley additive explanations that quantifies the contribution of each feature to a model's prediction across distributed data partitions. The computation distributes the sampling of feature coalitions across clients, with each node evaluating the model's output change when features are included or excluded from its local data. Results are aggregated to produce global feature importance scores. This technique is critical for auditing federated models for proxy discrimination, where seemingly neutral features like zip code encode protected attributes such as race or socioeconomic status.

A decentralized computation of the Area Under the Receiver Operating Characteristic Curve, a threshold-independent metric for binary classifiers. Rather than pooling raw prediction scores, each client computes the necessary components—such as sorted prediction scores and cumulative true positive and false positive rates—and shares only aggregated statistics. The global AUC is then reconstructed from these privacy-preserving aggregates. This metric is essential for evaluating diagnostic models across hospitals where different institutions may use different decision thresholds based on local clinical workflows.

A metric for quantifying shifts in the distribution of a model's input features between a reference period and a monitoring period, computed in a decentralized manner. Each client calculates the PSI for its local feature distributions using predefined bins, then shares only the per-bin population percentages. The central server aggregates these to detect data drift across the federated network. A PSI above 0.25 typically indicates significant distribution shift, triggering a need for model retraining or investigation into changing patient demographics at specific sites.