Federated Normalization

Computes standardized features (mean=0, std=1) across sites using secure aggregation of local sufficient statistics. Each site calculates local sums and squared sums, shares only these aggregates, and the central server derives global mean and variance without ever seeing raw values. This preserves the mathematical properties of z-score normalization while maintaining patient data locality.

Normalizes features to a fixed range (typically [0,1]) by discovering global minimum and maximum values through secure multi-party computation. Sites collaboratively determine the overall data range without revealing individual patient extremes. Critical for algorithms sensitive to bounded inputs, such as neural networks with sigmoid activation functions.

Aligns feature distributions by matching quantile values across decentralized nodes. Each site computes local quantile cut points, and a federated averaging protocol merges them into a global reference distribution. This technique is especially effective for non-IID clinical data where lab values follow institution-specific calibration curves.

Adapts batch normalization layers for decentralized training by tracking running means and variances locally, then synchronizing them during aggregation rounds. Prevents internal covariate shift across sites while avoiding the privacy leakage that would occur from sharing per-batch activation statistics directly.

Maps disparate local feature names and coding systems to a canonical global schema before normalization. Handles semantic equivalence (e.g., 'HbA1c' vs 'Hemoglobin A1c'), unit conversion (mg/dL to mmol/L), and categorical encoding mismatches. This preprocessing layer ensures normalization operates on semantically consistent inputs across all participating institutions.

Injects calibrated Laplace or Gaussian noise into shared normalization statistics to provide formal privacy bounds. The privacy budget (ε) controls the trade-off between statistical fidelity and patient protection. This ensures that even the aggregate normalization parameters cannot be reverse-engineered to infer individual-level information about any single patient in the federation.

Transforms local features to have a mean of 0 and standard deviation of 1. In a federated setting, sites share only aggregated statistics (sum, sum of squares, count) rather than raw values.

• Preserves the shape of the original distribution
• Sensitive to outliers in small local datasets
• Requires secure aggregation of local means and variances

Rescales features to a fixed range, typically [0, 1], using the global minimum and maximum values. Federated computation requires sharing only local extrema.

• Vulnerable to adversarial nodes reporting extreme values
• Useful when a bounded output range is required
• Often combined with clipping to handle outliers

Uses the median and interquartile range (IQR) instead of mean and standard deviation. This makes it resilient to outliers commonly found in clinical lab values.

• Ideal for skewed biomarker distributions
• Requires federated percentile computation
• Pairs well with differential privacy noise mechanisms

A technique that normalizes layer inputs within a neural network. In federated settings, local batch statistics can diverge due to non-IID data, causing convergence issues.

• FedBN variant keeps local BN parameters private
• Prevents catastrophic forgetting of site-specific distributions
• Critical for medical imaging models across different scanner vendors

Aligns the entire distribution of a feature across sites by mapping quantiles to a reference distribution. This handles non-linear distributional shifts that simple scaling cannot.

• Requires federated computation of empirical CDFs
• Common in genomics for normalizing gene expression data
• Preserves rank-order relationships between samples