Эксперимент: Channel-Noise Differentially Private Federated Optimizer (#363)

{ "worked": false, "confidence": 9, "verdict": "Built and numerically verified a toy channel-noise federated optimizer with clipping, disagreement-dependent Gaussian variance, damping, privacy proxy accounting, and contraction diagnostics. Geometric sensitivity sums matched theory, K95 saturation matched log(0.05)/log(rho) within 16% or better, and the stability boundary was exactly observed at |1-beta*lambda|=1. However, the predicted alpha^-1/2 privacy scaling failed because the sigma0 noise floor materially changes scaling, and the federated channel method produced only a 1.0% privacy-proxy reduction versus constant noise with essentially identical loss, so the promised effect was not demonstrated robustly.", "metrics": { "baseline": "Constant Gaussian noise: final loss 1.227809, cumulative privacy proxy 12112.01, mean variance 0.006400", "idea": "Channel-dependent noise: final loss 1.227809, cumulative privacy proxy 11991.52, approximately 1.0% lower proxy, mean variance 0.008283", "math_checks": "Geometric sums matched closed form to numerical precision; observed K95 steps were 5, 14, 29, 59 versus predictions 4.32, 13.43, 28.43, 58.40; stability diverged for beta>2 at lambda=1 as predicted; alpha-scaling reductions were 1.58, 2.92, 5.70 versus idealized 2, 4, 8." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py", "files": [ "experiment.py", "results.json", "run_output.txt" ], "limitations": "This is a small quadratic toy federated experiment rather than the proposed 10-client MNIST MLP. It uses an empirical clipped sensitivity proxy instead of a formal neighboring-dataset privacy accountant, does not estimate realized privacy loss, and tests only one main channel-noise parameter setting and one client/data configuration." }