#DifferentialPrivacy
🔏𝗥𝗲𝘃𝗶𝗲𝘄 𝗗𝗮𝘆 𝗧𝘂𝗲𝘀𝗱𝗮𝘆!📖

#DifferentialPrivacy is an increasingly popular, though controversial, approach. This week, @benrothke.bsky.social provides a Hall of Fame recommendation of @simson.net 's 𝘋𝘪𝘧𝘧𝘦𝘳𝘦𝘯𝘵𝘪𝘢𝘭 𝘗𝘳𝘪𝘷𝘢𝘤𝘺, from @mitpress.bsky.social.

👉 tinyurl.com/2s3jk4cn

#CybersecurityBooks #DP
October 14, 2025 at 1:35 PM Everybody can reply
2 reposts 4 likes
Concurrent composition theorems can improve adaptive differential privacy, researchers report. Read more: https://getnews.me/concurrent-composition-theorems-boost-adaptive-differential-privacy/ #differentialprivacy #adaptive
October 8, 2025 at 2:47 PM Everybody can reply
Researchers unveil the first differentially private algorithms for Wasserstein barycenters, delivering results on synthetic data, MNIST, and a U.S. population set. Read more: https://getnews.me/differentially-private-algorithms-for-wasserstein-barycenters/ #differentialprivacy #wasserstein
October 6, 2025 at 9:44 AM Everybody can reply
A denoising step after each DP‑SGD update restores low‑rank gradient structure, boosting sample efficiency when fine‑tuning RoBERTa on GLUE without compromising privacy. Read more: https://getnews.me/gradient-denoising-improves-efficiency-of-private-llm-fine-tuning/ #differentialprivacy #llm
October 3, 2025 at 3:50 AM Everybody can reply
A new scheme enables private estimation of black-box functions without sensitivity bounds, letting users balance data against query count via a parameter; privacy loss stays within ε‑δ. https://getnews.me/differential-privacy-scheme-for-black-box-function-estimation/ #differentialprivacy #blackbox
October 2, 2025 at 9:16 PM Everybody can reply
The paper classifies privacy attacks into three groups—anonymous re-identification, statistical reconstruction, and model inference—and notes defenses like noise perturbation. Read more: https://getnews.me/privacy-attacks-and-countermeasures-in-data-driven-systems/ #privacy #differentialprivacy
October 1, 2025 at 7:14 AM Everybody can reply
The Gaussian Pancake Mechanism (GPM) mimics the Gaussian mechanism but weakens differential‑privacy guarantees. 28 Sept 2025. Read more: https://getnews.me/backdoor-threat-to-differential-privacy-via-gaussian-pancake-mechanism/ #differentialprivacy #gaussianpancake #security
September 30, 2025 at 2:55 PM Everybody can reply
A new differentially‑private bootstrap provides valid confidence intervals, achieving nominal coverage on 2016 Canada Census data (reported 24 Sep 2025). https://getnews.me/new-differentially-private-bootstrap-improves-statistical-inference/ #differentialprivacy #bootstrap #statistics
September 27, 2025 at 5:59 AM Everybody can reply
Urania is a differential‑privacy framework for chatbot logs that uses private clustering, keyword extraction and histogram summarization, beating a non‑private Clio‑based baseline. Read more: https://getnews.me/urania-framework-enables-private-ai-chatbot-analysis/ #differentialprivacy #ai #chatbot
September 27, 2025 at 2:33 AM Everybody can reply
A new study shows LZ77’s global sensitivity is bounded by O(W^{2/3}·log n) and, when the window equals the input length, by Ω(n^{2/3}·log^{1/3} n). Presented at TCC 2025. Read more: https://getnews.me/differentially-private-compression-new-sensitivity-bounds-for-lz77/ #lz77 #differentialprivacy
September 27, 2025 at 12:47 AM Everybody can reply
New research shows the 2020 U.S. Census added more noise than needed, estimating that noise levels could be reduced by 15.08%‑24.82% while preserving privacy. Read more: https://getnews.me/2020-u-s-census-privacy-revealed-stronger-protection-than-expected/ #census #differentialprivacy #privacy
September 27, 2025 at 12:15 AM Everybody can reply
Monitoring aggregates deployment history to audit differential‑privacy guarantees, cutting required samples. Tests on Laplace and Gaussian mechanisms matched full‑audit accuracy. https://getnews.me/new-monitoring-method-tracks-differential-privacy-violations-over-time/ #differentialprivacy
September 26, 2025 at 9:20 PM Everybody can reply
A new study introduces Chernoff differential privacy tying the privacy budget ε to tighter error exponents for classifiers, with the paper submitted on 15 March 2024. Read more: https://getnews.me/chernoff-differential-privacy-boosts-adversarial-classification/ #chernoffdp #differentialprivacy
September 22, 2025 at 7:14 PM Everybody can reply
A new LDP algorithm cuts hypothesis‑selection queries to ~O(k^{3/2}) non‑adaptive queries for k candidate distributions, improving on prior quadratic methods. Submitted September 2025. https://getnews.me/query-efficient-private-hypothesis-selection-via-scheffe-graph/ #differentialprivacy #algorithm
September 22, 2025 at 4:58 PM Everybody can reply
LLMs can reverse‑engineer DP‑sanitized text, achieving 91.20% recovery on ChatGPT‑4o and 94.01% on Claude‑3.5 in black‑box attacks. Read more: https://getnews.me/llms-can-reconstruct-differentially-private-text-study-shows/ #largelanguagemodels #differentialprivacy #chatgpt4o
September 20, 2025 at 7:31 AM Everybody can reply
SynBench, a differential‑privacy text generation benchmark, tests datasets and finds quality drops sharply on complex data, even for large LLMs. It adds an attack to reveal privacy leaks. https://getnews.me/synbench-benchmark-shows-gaps-in-private-text-generation/ #differentialprivacy #llms
September 19, 2025 at 8:51 PM Everybody can reply
DPCheatSheet is a web tool that helps novices add differential privacy to LLM‑generated code. In a study of six novices and three DP experts, users flagged more errors. https://getnews.me/dpcheatsheet-helps-novice-developers-implement-differential-privacy-with-llms/ #differentialprivacy #llm
September 18, 2025 at 8:36 AM Everybody can reply
Google released VaultGemma, a 1‑billion‑parameter private LLM that matches non‑private models on benchmarks like MMLU. Code is on Hugging Face. Read more: https://getnews.me/google-unveils-vaultgemma-a-high-performance-private-ai-model/ #vaultgemma #differentialprivacy #google
September 18, 2025 at 2:42 AM Everybody can reply
FedRP combines random projection with ADMM to give ε,δ differential‑privacy and drastically lower communication traffic. The paper was submitted on 12 Sep 2025. Read more: https://getnews.me/fedrp-efficient-private-federated-learning-via-random-projection/ #federatedlearning #differentialprivacy
September 17, 2025 at 2:50 AM Everybody can reply
Google released VaultGemma 1B, a small but fully DP-trained open-weight LLM. It trades some accuracy for strong privacy (ε≤2, δ≤1.1e-10, at 1,024-token sequences), shows no detectable memorization on standard probes, and is built with conventional transformer parts. #differentialprivacy #ai #gemma
September 14, 2025 at 10:58 AM Everybody can reply
An interesting #AI model from Google that could be very useful for #Biomedical & #Health applications, where you need to deal no only with #PII but also #PHI:

"VaultGemma: The world's most capable differentially private #LLM"

research.google/blog/vaultge...

#DifferentialPrivacy
VaultGemma: The world's most capable differentially private LLM
research.google
September 12, 2025 at 6:24 PM Everybody can reply