← All AI/ML open problems

Privacy & encrypted learning · Learning dynamics & scaling / Open question

Learning Useful Features Under Privacy Noise

Clipping and privacy noise change which features can emerge, not just the final parameter variance. Ambient dimension and effective signal dimension can have different costs. Efficiency payoff: Identify representations and width/update schedules that spend scarce private examples and compute on useful signal.

OpenEditorially scopedFramework needs definitionCompiled 2026-11-20

Seed catalog · Showing the bundled record.

Problem statement

For X∼N(0,Id)X\sim N(0,I_d) and Pr⁡(Y=1∣X)=sigmoid⁡(k−1/2∑j=1kajtanh⁡(uj⊤X))\Pr(Y=1\mid X)=\operatorname{sigmoid}(k^{-1/2}\sum_{j=1}^k a_j\tanh(u_j^\top X)), with orthonormal unknown uju_j and ∣aj∣≤1|a_j|\leq1, train both layers of a width-mm network using clipped per-example gradients and Gaussian noise. Determine the excess logistic-risk frontier at sample size nn, TT steps and total (εDP,δDP)(\varepsilon_{\rm DP},\delta_{\rm DP}) privacy, comparing learned features with a same-budget frozen-feature learner.

Assumptions and scope

  • Central example-replacement DP means Pr(A(D) in E)<=exp(epsilon_DP)Pr(A(D') in E)+delta_DP for every event and every neighboring pair, not only likely teacher-generated datasets. Take epsilon_DP in (0,1] and delta_DP<n^(-2); report dependence on both.
  • Fix subsampling, clipping thresholds, initialization, parameter bounds, learning rates, noise and accountant. Compose all training, private hyperparameter selection and any private projection discovery. Public pretraining must be independent of the protected data and its information/cost stated.
  • Risk is expected fresh-example logistic loss minus the teacher's Bayes loss. Charge per-example gradient clipping, total arithmetic and optimizer memory; a small stationarity residual or a large signal-to-noise ratio alone is not a population-risk guarantee.

What would settle it?

A finite-data, finite-training risk upper bound with a strict resource-matched benefit or a privacy-induced obstruction; separate representation, clipping, optimization and privacy terms and provide a lower-bound comparison.

Selected for a theoretical or computational target. A fitted scaling curve or improved benchmark score alone does not establish the stated learning-theoretic guarantee.

Status and evidence

DP-SGD, feature-centric private CNN theory and DP-muP spectral scaling already exist. This orthonormal tanh-teacher benchmark asks for a joint risk/compute/privacy law, not the first private feature-learning theorem; its precise optimizer and scaling regime remain to be specified. This is an editorial, source-motivated research target. The preliminary review through 2026-11-20 does not certify novelty or exclude an unlocated resolution.

Preliminary source check; not an expert-verified open status or an independent audit of proposed proofs. The compilation date is not a literature-review cutoff.

Sources and further reading

  1. Abadi et al. — Deep Learning with Differential Privacy ↗partial theory · source checked 2026-11-20

    Primary abstract checked: clipped/noisy neural training and refined composition are established. Privacy accounting alone does not prove feature learning or an optimal risk/compute trade-off.

  2. Ding et al. — Understanding Private Learning From Feature Perspective ↗partial theory · source checked 2026-11-20

    Primary search abstract checked: two-layer polynomial-ReLU CNN feature-signal learning and memorization under noisy gradient descent. Private feature learning is not an unexplored existence question.

  3. Xu and Chen — Differential Privacy in Two-Layer Networks: How DP-SGD Harms Fairness and Robustness ↗partial theory · source checked 2026-11-20

    Primary search abstract checked: feature-to-noise-dependent loss and robustness/subpopulation effects, including limits of public-pretraining transfer. The proposed target must retain this existing feature-centric theory.

  4. Wang, Bu and Zhang — Towards Hyperparameter Transfer for Differentially Private Optimization ↗partial theory · source checked 2026-11-20

    Primary abstract checked: DP-muP derives spectral/SNR scaling rules for DP-SGD, Adam and Muon and tests width transfer. A uniform finite-training risk law is a separate claim.