Differential Privacy Feature Processing for Vertical Federated Learning

The patent introduces differential privacy as an intermediary mechanism between data holders and data users. Instead of direct data access, the system uses noisy aggregated statistics from multiple parties to perform feature processing. This mediator layer enables collaborative feature selection and coding while preventing any single party from exposing their raw data, thus resolving the contradiction between processing capability and privacy security.

The patent segments the feature processing task into separate contributions from different parties. Each party independently computes local statistics on their data segments, and these segmented results are aggregated through secure multiparty computation. This segmentation allows feature processing to proceed across parties without requiring centralization of raw data, maintaining both productivity and privacy.

The patent creates encrypted copies of label information that can be used for WoE and IV calculations without exposing the actual labels. Through homomorphic encryption and differential privacy, the system generates computational copies that preserve the statistical properties needed for accurate feature evaluation while being mathematically impossible to reverse to the original sensitive data, thus maintaining measurement precision without privacy loss.

The patent transforms the label data from its original sensitive form into differential privacy-protected statistical parameters. Instead of sharing raw labels, parties share noisy aggregated counts and statistics that are sufficient for calculating WoE and IV indices. This parameter transformation maintains the accuracy needed for feature evaluation while fundamentally changing the data representation to eliminate privacy risks.

The patent applies differential privacy noise at the appropriate granularity level - adding noise to aggregated statistics rather than individual records. This partial application of noise provides sufficient privacy protection for the security requirement while minimizing the impact on calculation precision. The noise level is calibrated to be excessive only where needed for privacy, not excessively applied throughout the entire computation.

The patent预先 cushions the impact of noise on precision by using robust statistical methods in the WoE and IV calculations. The feature evaluation algorithms are designed to be resilient to differential privacy noise, and the system pre-calculates multiple noisy estimates that can be aggregated to reduce variance. This beforehand cushioning ensures that while individual noisy measurements have reduced precision, the final feature indices maintain sufficient accuracy for practical use.