SegFL: Federated Behavioural Segmentation
A privacy-preserving, client-aware federated learning framework designed to perform behavioural segmentation across multi-tenant systems without sharing raw data.
Key Highlights
- Tenant Adapter Layer (TAL)
- Vectorized DP-SGD optimization
- Post-hoc surrogate explainability
The Challenge
Centralized behavioural modelling compromises user privacy and struggles with data heterogeneity across different organizations. Standard federated learning models face severe bottlenecks when dealing with mismatched feature schemas and non-IID data distributions.
Our Approach
We engineered a novel Tenant Adapter Layer (TAL) to align heterogeneous database schemas locally, coupled with a deep bottleneck autoencoder. The system utilizes vectorized DP-SGD for mathematical privacy guarantees and unsupervised clustering for segmentation.
The Result
The framework successfully generated high-quality customer personas with a clustering Silhouette score of 0.800, matching centralized performance while strictly bounding privacy leakage. The solution was deployed via an interactive Streamlit dashboard for actionable insights.
Research & Automation Impact
Ready to build something similar?
Let's discuss how we can transform your digital presence.