Better noise factorization for private machine learning

Differentially private training uses correlated noise to improve model utility while preserving privacy guarantees. Existing methods like BISR work well with large noise buffers but degrade when memory is constrained. γ-BIFR generalizes prior approaches, achieving significantly better RMSE and error bounds across all bandwidth regimes—from memory-limited to asymptotically optimal settings. The method is analyzed theoretically for multi-epoch training and provides tighter guarantees for multi-participation scenarios.