Can private prompt optimization avoid token-by-token brittleness? DP-ES says yes, with noisy evolution strategies.
trending23
01 02 Paper derives closed-form MIA bounds for random-allocation DP-SGD, so you can tune noise without the usual accountant pilgrimage.03 Analyzes DP for deterministic GD on a Gaussian-perturbed objective, with finite-iterate privacy via a diffeomorphism proof.04 BetweenCut: Private Heavy-Node Classification with Doubly Logarithmic Error in Tree Height arxiv.orgBetweenCut does private heavy-node classification on trees with O(log log h) error, retiring the usual log h tax.05 Federated Bayesian Surveillance of Mechanical Thrombectomy Adverse Events: A Population Risk Layer for Surgical Digital Twins arxiv.orgFederated Gamma-Poisson risk layers for thrombectomy twins, with RDP natural-parameter updates instead of yet another DP FedAvg.06 DP tabular foundation models are here, with PrivTab baking the privacy mechanism into the model instead of bolting it on.07 Learns a DP-safe linear mix of public datasets before private fine-tuning, instead of picking one “best” corpus.08 Beyond Monolithic Perturbation: Heterogeneous Mechanism Design for Multi-Attribute Metric Differential Privacy arxiv.orgMetric-DP for mixed-attribute records gets dependency-aware grouping and per-attribute mechanisms, instead of one Laplace for all.09 Paper unifies DP accountants, then shows zCDP throws away curve shape, which is apparently where the useful bits lived.10 Combining Homomorphic Encryption and Differential Privacy in Federated Learning for Model Inspection and Availability arxiv.orgPaper glues HE training to DP for model inspection/release in FL, then claims better FEMNIST utility/privacy than DP-only.11 Paper on quantum two-party DP for Hamming distance, with O(n) communication and constant expected error, because classical wasn't enough.12 OpenDP's discrete Laplace sampler had periodic bias, traced to a buggy bernoulli_exp1 rational path and fixed with exact arithmetic.13 What privacy guarantee actually survives when multiple labs, agents, and side channels all get a vote?14 PrivCert formalizes when a DP text model can certify support, or admit it has none, with tradeoffs for honesty versus privacy.15 Privacy-Friendly Cohort Determination: Sealed, CSP-Independent In-Browser ML Inference of Professional Segments for Identity-Less Advertising arxiv.orgA paper on browser-side cohort inference for ads, with local DP and no cross-site IDs, because tracking wasn't baroque enough.16 Paper on certified robustness for ViTs, showing PixelDP after patch embedding can drop to chance before patching it with spectral norms.17 Federated learning with TEEs, policy-bound uploads, and audit logs to make central DP externally verifiable, not just promised.18 Is it safe to lie? Revisiting Incoercible Multi-Party Computation in the Universal Composability Framework eprint.iacr.orgUC incoercibility for MPC, plus hardware tokens and Γ-deniability, because even lying now needs a formal model.19 Paper on ROGA, a resizable oblivious group-by scheme with Rust and Binsec/Rel proofs, because ORAM was not slow enough.20 Only Pay What You Must Spend: On-Demand Privacy Budget Payment for Differentially Private RAG arxiv.orgCan RAG spend privacy budget only when it actually needs private data? SparsePay-RAG claims to do that with clustering and DP decoding.21 DP inputs hide graph shift operators when the filter has the right zeros, with Gaussian guarantees and a CRB bound.22 Pattern-level DP for complex event processing shifts noise from raw streams to detected patterns, for once.23 Fairness beats security-with-abort for DP outputs, if you enjoy proving aborts can leak just enough to matter.