Federated SMOTE

Federated SMOTE generates synthetic minority class samples by performing interpolation between feature vectors that remain distributed across local nodes. Instead of centralizing data, each institution computes local nearest neighbors and shares only anonymized statistical summaries or securely aggregated interpolation directions. This preserves the core SMOTE mechanism—creating new samples along line segments connecting existing minority points—while ensuring raw patient records never leave their originating hospital.

To prevent membership inference attacks that could reveal whether a specific patient's record contributed to synthetic samples, Federated SMOTE integrates calibrated noise injection into the interpolation process. By applying differential privacy mechanisms during neighbor selection and sample generation, the algorithm provides a provable mathematical guarantee that individual patient contributions remain indistinguishable. This is critical for HIPAA and GDPR compliance in multi-institutional clinical research networks.

Clinical data across hospitals is inherently non-independent and identically distributed (non-IID) —a rural clinic may see different disease prevalence than an urban research center. Federated SMOTE addresses this by allowing site-specific oversampling ratios tailored to local class imbalances while maintaining global model coherence. Each node can adjust the amount of synthetic generation based on its own minority class scarcity without distorting the collaborative learning objective.

Rather than transmitting generated synthetic samples directly, Federated SMOTE employs secure multi-party computation (SMPC) or homomorphic encryption to aggregate interpolation parameters across nodes. The global model receives only encrypted gradient updates derived from augmented local datasets, ensuring that even the synthetic samples remain opaque to other participants. This defense-in-depth approach protects against both raw data exposure and synthetic data reconstruction attacks.

Transmitting full synthetic datasets across federated networks would create prohibitive bandwidth overhead. Federated SMOTE optimizes this by sharing only compact interpolation coefficients or latent space representations rather than high-dimensional clinical feature vectors. Techniques like gradient compression and sparse vector transmission reduce communication costs by orders of magnitude, making the approach viable even across hospitals with limited network infrastructure.

Federated SMOTE operates as a preprocessing step within standard Federated Averaging (FedAvg) pipelines. Before each round of local model training, minority class samples are augmented using the privacy-preserving SMOTE variant. The augmented local datasets then train models whose weight updates are aggregated by the central server. This modular design allows Federated SMOTE to enhance existing federated learning deployments without requiring architectural overhauls.

The foundational algorithm that Federated SMOTE extends. SMOTE generates synthetic samples for underrepresented classes by interpolating between existing minority data points in feature space.

• Core Mechanism: Selects a minority sample, identifies its k-nearest minority neighbors, and creates new points along the line segments connecting them
• Key Limitation: Requires centralized access to all data, making it incompatible with privacy regulations like HIPAA
• Variants: Borderline-SMOTE focuses on samples near the decision boundary; ADASYN adapts sampling density based on learning difficulty

A decentralized generative adversarial network where the discriminator and generator are trained across multiple institutions without sharing underlying patient data. Unlike Federated SMOTE's interpolation approach, Federated GANs learn to model the full data distribution.

• Architecture: Local generators produce synthetic data; a federated discriminator provides privacy-compliant feedback
• Use Case: Generating entire synthetic patient cohorts for rare disease research
• Trade-off: Higher computational cost than Federated SMOTE but produces more diverse synthetic samples

A mathematical framework that injects calibrated noise into data or model updates to provide provable guarantees against patient re-identification. Often paired with Federated SMOTE to bound privacy leakage from synthetic samples.

• Privacy Budget (ε): Quantifies the maximum information leakage allowed; lower epsilon means stronger privacy
• Integration: Noise can be added during neighbor selection or interpolation in Federated SMOTE
• Gaussian Mechanism: Adds noise proportional to the sensitivity of the SMOTE computation

The challenge of managing statistical heterogeneity across clinical sites where patient populations differ significantly. Federated SMOTE must account for varying class imbalance ratios at each institution.

• Label Distribution Skew: Hospital A may have 30% positive cases while Hospital B has only 5%
• Feature Distribution Skew: Different patient demographics create divergent feature spaces
• Mitigation: Federated SMOTE can apply site-specific oversampling ratios rather than a global strategy

A quantitative measure of how well synthetic samples preserve the statistical relationships and predictive performance of real data. Critical for validating Federated SMOTE outputs without accessing raw patient records.

• Train-Synthetic-Test-Real (TSTR): Model trained on synthetic data, evaluated on real holdout sets
• Propensity Score Matching: Measures how distinguishable synthetic samples are from real ones
• Feature Correlation Preservation: Ensures clinical relationships between variables remain intact

Decentralized algorithms that collaboratively learn to fill in missing clinical data values across institutions. Often used as a preprocessing step before Federated SMOTE to handle incomplete minority class records.

• Missing Not At Random (MNAR): The most challenging pattern where missingness relates to unobserved values
• Multiple Imputation: Generates several plausible values to capture uncertainty
• Synergy: Complete records enable more accurate k-nearest neighbor selection in Federated SMOTE