Federated SHAP

Computes Shapley values without centralizing raw patient data. Each client site calculates local explanations on its own siloed dataset, and only the aggregated statistical summaries are shared with the central server. This ensures compliance with HIPAA and GDPR mandates while still providing a global view of which biomarkers drive predictions.

• No raw data leaves the hospital firewall
• Local SHAP computations use the site's own data distribution
• Aggregated attributions reflect the full federated population

Reconstructs a unified explanation of the federated model's behavior across all participating institutions. By aggregating local Shapley values, Federated SHAP reveals which features are consistently important globally, even when individual sites have non-IID data distributions. This allows a CTO to understand if the model relies on clinically valid biomarkers or spurious correlations.

• Identifies features with high global importance
• Detects heterogeneity in feature effects across sites
• Provides a single pane of glass for model auditing

Beyond a single global explanation, Federated SHAP can preserve local explanation fidelity. The architecture allows analysts to inspect how feature importance varies between a community hospital and a tertiary research center. This is critical for detecting dataset shift and understanding why a model performs differently on distinct subpopulations.

• Compare feature rankings between Client A and Client B
• Identify site-specific outlier explanations
• Validate that the model uses appropriate context for each demographic

Transmitting full local explanation matrices is bandwidth-prohibitive. Federated SHAP implementations often use quantized histograms or sufficient statistics to compress Shapley value distributions before transmission. This reduces the communication overhead to a fraction of the raw data size while preserving the statistical integrity of the global feature importance ranking.

• Uses compressed gradients of explanation distributions
• Compatible with secure aggregation protocols
• Minimizes the carbon footprint of cross-silo training

Even aggregated Shapley values can leak membership information. Federated SHAP can be hardened by injecting calibrated Gaussian noise into the local explanation statistics before aggregation. This provides a formal ($\epsilon$, $\delta$)-differential privacy guarantee, ensuring that the presence of any single patient in a local cohort cannot be inferred from the published global feature importance report.

• Applies DP-SGD principles to explanation vectors
• Balances privacy budget against explanation accuracy
• Essential for publishing research on federated biomarker models

Exact Shapley value computation scales exponentially with feature count. Federated SHAP leverages KernelSHAP or TreeSHAP approximations locally, with the federated layer aggregating these estimates. For high-dimensional genomic data, feature grouping or hierarchical Shapley methods are employed to keep local computation feasible on standard hospital IT infrastructure.

• Uses Monte Carlo sampling of feature coalitions
• Leverages model-specific optimizations (e.g., TreeSHAP for XGBoost)
• Avoids the $O(2^n)$ combinatorial explosion

The application of SHAP values and attention visualization to decode clinical AI decision-making. In regulated diagnostic contexts, explainability is not optional—it is a regulatory prerequisite for FDA and CE mark submission.

• Generates patient-specific feature attributions for each prediction
• Enables clinicians to verify that models rely on clinically relevant biomarkers
• Supports audit trails required by 21 CFR Part 11 and EU MDR
• Complements federated SHAP by providing the local explanation framework that federated SHAP aggregates across sites

A cryptographic protocol that allows a central server to compute the sum of model updates from multiple clients while ensuring the server cannot inspect any individual client's contribution in plaintext.

• Uses secret sharing or pairwise masking to blind individual updates
• Critical for federated SHAP because local SHAP values can leak distributional information about private data
• Ensures that only the aggregated global explanation is revealed, not per-patient feature contributions
• Often combined with differential privacy for formal guarantees

A mathematical framework that injects calibrated statistical noise into data or model updates to provide a provable guarantee that the presence or absence of any single individual's record cannot be inferred.

• Parameterized by epsilon (ε) — lower values mean stronger privacy
• Applied to federated SHAP outputs to prevent membership inference from explanation vectors
• The Gaussian mechanism and Laplace mechanism are common noise-injection methods
• Balances the privacy-utility tradeoff: too much noise degrades explanation fidelity

A security vulnerability where raw model gradients shared during distributed training can be analytically inverted by a malicious server to reconstruct the private input data that generated them.

• Deep Leakage from Gradients (DLG) and Inverting Gradients are canonical attacks
• Federated SHAP faces analogous risks: local SHAP values encode information about the underlying data distribution
• Mitigation requires combining federated SHAP with secure aggregation or homomorphic encryption
• Demonstrates why naive federation of explainability methods is insufficient without cryptographic safeguards

A data distribution characteristic in federated settings where local client datasets are statistically heterogeneous and do not represent the overall population distribution.

• Manifests as label skew, feature skew, or concept drift across sites
• Poses a fundamental challenge for federated SHAP: global feature importance may not reflect any single site's reality
• Requires careful interpretation — a feature that is globally important may be irrelevant at a specific hospital
• FedProx and personalized federated learning are algorithmic responses to this heterogeneity

A federated learning topology designed for a small number of reliable institutional clients, such as hospitals, that hold large, curated datasets and participate in every training round.

• The primary deployment context for federated SHAP in healthcare
• Assumes stateful clients with dedicated compute infrastructure
• Enables complex per-site computation like exact SHAP value calculation using KernelSHAP or TreeSHAP
• Contrasts with cross-device FL, where intermittent availability makes explanation aggregation more challenging