Discover01 02 03 See all in Privacy →
Digestby Monstream · Every Thursday, 09:00 UTC
Privacy-Preserving ML Papers
Research on training and running models without exposing data: differential privacy, federated learning, secure inference and multi-party computation for ML.
78Subscribers
1Results so far
Oct 2026Created
1d agoLast update
Latest issue
Oct 6 · 3 of 12 shownThis issue has 11 items: differential privacy theory and practice, federated learning, and homomorphic encryption for language models. Most worth reading are DP-ES, which stabilises private prompt optimization, and AR-HE, which runs the full decoding loop under encryption on the server. TranScope, a hardware membership-inference attack, was left out because it matches no secondary topic.
Private online learning and prediction for Littlestone classes
Shows that private online learning and private online prediction are separated for classes of finite Littlestone dimension. Every (ε,δ)-private online learner has a stream with expected mistakes of at least Ω((d/ε)·log(T)^{2/3}), while a jointly private predictor makes at most 2^{2^{cd²}}ε⁻² log²(2/(εδ)) expected mistakes independent of T.
Differentially Private Mixing of Public Datasets Improves Private Learning
Introduces a pipeline that privately learns the mixture of several public datasets to pretrain on before DP finetuning, using a privately learned low-dimensional linear model. On NIH ChestX-ray14 it improved macro AUC by up to 0.037, with +22.8% relative AUC on Cardiomegaly at ε=1. On ENRON it cut test perplexity by 16% against baseline mixtures.
🔒9 more in this issueSubscribe to see everything this stream finds, in your feed and by email. It’s free; one run serves every subscriber.
More in Privacy
Streams that already watch this field. Follow one as it is — it costs nothing extra.
RadarDaily batch, 08:00 UTC
Everything in fully homomorphic encryption: new papers and benchmarks, libraries and releases, new projects and startups, funding, products, standards, talks and events, attacks and incidents — across research, industry and onchain FHE.
84 subscribers
RadarDaily batch, 09:00 UTC
Everything around FHERMA, the FHE challenge platform by Fair Math: new challenges and prize pools, results and winners, solution write-ups, Polycircuit and OpenFHE-rs releases, partnerships and mentions.
80 subscribers
RadarAs found
Every new release of the homomorphic encryption libraries, compilers and SDKs — TFHE-rs, Concrete, fhEVM, OpenFHE, SEAL, HEIR, Lattigo, Poulpy, CryptoLab HEaaN and enVector, Desilo, IBM HElayers, Apple Swift HE, TenSEAL and GPU libraries — with what changed.
80 subscribers
Can’t find what you need? Describe it in a sentence and Monstream builds the stream for you.Create your own →