Federated Meta-Learning Algorithm for Cross-Enterprise Collaborative Business Intelligence Without Data Sharing

Abstract

Introduce FedMeta-BI – a Federated Meta-Learning Framework where multiple competing companies collaborate to train Business Intelligence (BI) models without exposing any proprietary raw data. In our approach, FedMeta-BI integrates Model Agnostic Meta-Learning (MAML) with a secure federated optimisation algorithm utilising differentially private stochastic gradient descent (DP-SGD, ε = 1.0) and selective homomorphic encryption (HE) for gradient aggregation. Our method uses a cross-enterprise knowledge graph to incorporate structural priors across multiple domains for better few-shot adaptation and transferability to new enterprise settings. Evaluate FedMeta-BI on five anonymised enterprise datasets drawn from Finance, Retail, Healthcare, Logistics, and Manufacturing domains having more than 20 million entries in total. In our experimental setup, achieve a Macro-F1 of 0.891, outperforming FedAvg by 8.5%, MAML by 13.7% and local training by 25.2%. In our differentially-private setup with strict DP guarantees (ε=1.0), obtain an accuracy drop of just 2.3% compared to non-private version. The membership inference attack AUC drops to 0.512, which is no better than random chance. Communication costs reduced by 56% compared to FedAvg while maintaining similar model performance.