Secure Ensemble Training Using Heterogeneous Private ML Models

The system segments the ensemble model into multiple local models distributed across different data owners. Each data owner trains and maintains their own local model, which processes queries locally without exposing raw data. The service provider receives only aggregated predictions from these segmented local models, enabling comprehensive responses while preserving individual data privacy through spatial and functional segmentation of the modeling process.

The service provider acts as an intermediary that coordinates the meta-model training process without requiring direct peer-to-peer communication between data owners. The service provider collects local model predictions, aggregates them, and facilitates the training of the meta-model that learns to combine these predictions. This intermediary role simplifies coordination complexity by centralizing the aggregation logic while still enabling collaborative improvement of model accuracy through the meta-model's ability to weigh and combine multiple local model outputs.

The system implements partial action by having each data owner's local model process only the subset of queries relevant to their data domain, rather than all queries. The meta-model selectively aggregates predictions from local models based on query relevance and model confidence, avoiding unnecessary computation from all data owners for every query. This partial aggregation strategy maintains response accuracy by involving only the necessary subset of local models while reducing overall computational overhead and communication costs compared to full aggregation of all local model predictions.