Decentralized federated distillation for privacy-preserving cross-league basketball data collaboration

Abstract

Abstract Cross-league basketball analytics promises richer, more transferable performance models, yet competitive sensitivities and data-protection regulations make raw data sharing across leagues impractical. We propose a decentralized federated distillation framework that lets multiple basketball leagues co-train predictive models without centralizing data and without depending on a trusted aggregator. Each league node trains a locally chosen model on its anonymized aggregate game-level statistics and exchanges only temperature-scaled soft predictions with neighbors over a sparse peer-to-peer graph. To address re-identification threats and cross-league feature-space mismatch, the pipeline pairs an ε-differential-privacy Laplace mechanism applied directly to the released soft predictions—with explicit Rényi-DP composition across rounds—with k-anonymity for quasi-identifier coarsening and a Wasserstein optimal-transport projection that aligns league-specific feature spaces into a shared 64-dimensional representation. We establish convergence guarantees for federated distillation over decentralized communication graphs under non-convex objectives and heterogeneous data distributions, deriving an explicit bound that exposes the joint role of network spectral gap, distillation approximation error, transport-alignment error, and data heterogeneity. On a four-league dataset spanning the NBA, CBA, EuroLeague, and KBL—33,048 games in total—the proposed method attains 78.4% game-outcome accuracy, only 1.8 points behind a centralized oracle, while cutting communication overhead by more than 98% relative to parameter-averaging alternatives and preserving formal differential-privacy guarantees. Ablation studies confirm that feature alignment and adaptive temperature scheduling are both indispensable, and the sparse custom topology balances convergence speed against bandwidth efficiency.