differential privacy
Differential privacy is a way to analyze data while mathematically limiting how much information leaks about any single person. It works by adding carefully calibrated random noise to the results, so that the outcome hardly changes whether a specific individual's data is included or not. This protects privacy even if an attacker has extra information. The goal is to get useful insights without exposing private details.
The papers explore making differential privacy more practical across diverse settings. They show that tailored public pretraining boosts accuracy under privacy, that long-context tasks can be handled by splitting inputs and using local denoising, and that tighter composition analysis allows less noise for given privacy guarantees. They also achieve optimal regret in online learning, prove that private distribution learning is always possible when non-private learning is, and improve the sample complexity of private PAC learning to near-optimal levels. Overall, the advances focus on reducing the utility cost of privacy and broadening the range of problems where differential privacy can be effectively applied.
- Papers
- 20
- Released code
- 1
- First seen
- Sept 2026
- Latest
- Nov 2026
20 papers in the last two months, against 0 in the two before.
Start here
Three papers to read first, in the order they came out- 1Filtering atypical records lets differential privacy noise match typical data sensitivity
Propose-filter-release privately filters data to calibrate differential privacy noise to typical, rather than worst-case, record influence.
Oct 2026 - 2Privacy-protecting averages can lose accuracy when data variation is unknown
Private Gaussian mean estimation cannot achieve dimension-free error when the covariance matrix is unknown.
Nov 2026 - 3Privacy-preserving watermarks for AI text allow aggregate tracking without individual surveillance
Privacy and provenance for language model watermarks can be reconciled by shifting from individual tracking to aggregate frequency estimation.
Nov 2026
How fast it is moving
Often studied with
synthetic data generationCreating training examples artificially, often using an LLM, to overcome real-data scarcity or privacy constraints.3 togethermemorizationMechanisms by which a model stores and reproduces exact training data, especially sensitive information, rather than learning generalizable patterns.2 togetherlarge language modelA neural network trained on massive text corpora to generate coherent text and perform diverse language tasks.2 togetherWho is working on it
The papers
Most central to this idea first, not most recent.
- Top Universitycs.LG
Privacy-protecting averages can lose accuracy when data variation is unknown
Private Gaussian mean estimation cannot achieve dimension-free error when the covariance matrix is unknown.
École polytechnique, UC Berkeley · Nov 2026
- Top Universitycs.CR
Privacy-preserving watermarks for AI text allow aggregate tracking without individual surveillance
Privacy and provenance for language model watermarks can be reconciled by shifting from individual tracking to aggregate frequency estimation.
Stanford University · Nov 2026
- Top Universitycs.LG
Filtering atypical records lets differential privacy noise match typical data sensitivity
Propose-filter-release privately filters data to calibrate differential privacy noise to typical, rather than worst-case, record influence.
Hebrew University of Jerusalem, Bar-Ilan University · Oct 2026
- Big Techcs.LG
Every learnable distribution class is also differentially privately learnable
Every distribution class learnable with arbitrarily small error in total variation distance is also learnable under differential privacy.
University of British Columbia, McMaster University · Oct 2026
- Big Techcs.CR
Data privacy tools calculate more accurate totals by considering outliers’ direction
A new method privately selects a clipping bound for DP vector sums by optimizing residual coherence and outperforms existing approaches.
Google Research · Oct 2026
- Big Techcs.CR
Guided diffusion enables trustworthy synthetic text with formal privacy guarantees
Guided Diffusion steers diffusion language models toward target semantic embeddings to generate synthetic text with differential privacy and high utility.
NVIDIA, Carnegie Mellon University · Oct 2026
- Big Techcs.LG
Adding dummy points hides batch sizes and preserves privacy under truncation
Adding dummy points to batches restores the privacy guarantee of DP-SGD when batch sizes are observable or truncated.
Apple · Nov 2026
- Top Universitycs.LG
Empirical denoiser enables privacy accounting for diffusion sampling
Privacy loss in diffusion sampling concentrates in the critical window where semantics form and can be certified and reduced by subsampling, early stopping, and hybrid public denoisers.
University of British Columbia, Stanford University · Oct 2026
- Big Techcs.LG
AI training’s privacy losses resemble a bell curve when individually small
Privacy loss of composing subsampled differentially private algorithms is approximately Gaussian, after truncating rare large losses, with small logarithmic shifts.
Apple, The Hebrew University of Jerusalem · Oct 2026
- Top Universitycs.LG
Differentially private online prediction achieves regret proportional to square root of best expert's loss
A private online prediction algorithm achieves regret O~(sqrt(L*)), matching nonprivate performance up to polylog factors when inputs are oblivious.
Guangzhou HKUST Fok Ying Tung Research Institute, HKUST · Oct 2026
- Big Techcs.LG
Nearly linear sample complexity suffices for private learning of Littlestone classes
Approximate differentially private PAC learning of binary Littlestone classes needs O(d log d) examples, nearly optimal up to a logarithmic factor.
Technion - Israel Institute of Technology, Google Research · Oct 2026
- Chinese Techcs.AI
Memory-guided framework brings differential privacy to million-token LLM inference
DP-MIND enables private inference on million-token contexts by perturbing chunks, propagating memory, and local denoising.
Zhejiang University, Ant Group · Oct 2026
- Big Techcs.LG
Embedding-free framework enables private topic discovery in LLM conversations
CALLIOPE produces differentially private topic hierarchies from text using LLMs directly, beating embedding-based methods in the high privacy regime.
Google Research, Google DeepMind · Oct 2026
- Big Techcs.CRcode
New methods tighten privacy accounting for DP synthetic data pipelines
GDP and PLD analyses of bounded-range mechanisms enable tighter end-to-end accounting for mixed DP synthetic data pipelines, improving utility at the same privacy level.
UCL, SAS · Oct 2026
- Top Universitycs.LG
PrivTab embeds differential privacy into a tabular foundation model for fast private classification
PrivTab delivers differentially private tabular classification in one forward pass, outperforming private linear and neural baselines under moderate-to-strong privacy while fitting 10,000 times faster.
University of Helsinki, University of Cambridge · Oct 2026
- Top Universitycs.LG
Privately learning public data mixtures improves differentially private training
A pipeline that privately learns the mixture of public datasets for pretraining improves private downstream task performance.
University of Toronto, Vector Institute · Oct 2026
- Big Techcs.CR
Privacy-aware rewriting offers best privacy-capability trade-off for LLMs on sensitive data
Privacy-aware rewriting of synthetic data achieves a more favorable privacy-capability trade-off than direct finetuning or DP-SGD.
Johns Hopkins University, OpenAI · Oct 2026
- Academiccs.CR
Continually updated AI models reveal when data was used in training
A black-box attack infers whether and when individual training samples entered or left a continually updated model, beating prior methods by reading confidence patterns across snapshots.
Stevens Institute of Technology, University of Connecticut · Nov 2026
- Big Techcs.LG
Near-linear time random Fourier features with strong concentration guarantees
Random Fourier features for shift-invariant kernels can be computed in near-linear time with near-subgaussian concentration.
Microsoft Research, University of Wisconsin-Madison · Oct 2026
- Big Techcs.CL
Tokenwise reference penalties curb privacy memorization during language-model fine-tuning
TRAP penalty reduces memorization of sensitive data in fine-tuned language models with little utility loss.
Carnegie Mellon University, NVIDIA · Sept 2026
Concepts are extracted from each paper and reused across the corpus, so this page grows on its own as the desk reads.