NDSS 2025 – MingledPie: A Cluster Mingling Approach For Mitigating Preference Profiling In CFL

This post highlights an NDSS 2025 session paper, MingledPie, a privacy-enhanced clustered federated learning framework that mingles client types across clusters to thwart server-side preference profiling; it introduces indistinguishable cluster identities and a linear-equation-based reconstruction to recover accurate models despite mixed members, with theoretical analysis and experiments on six datasets showing defense effectiveness (reported 69.4% average) and minimal model accuracy loss (0.02%–3.00%).