StraightDP: Geometry-Aware Differential Privacy for Rectified-Flow Transformers
Not on the map. Why?
Abstract (arXiv)
Differentially private (DP) training of text-conditioned generative models suffers a utility cliff at strong privacy. We revisit this problem through the geometry of rectified flows: along the straight interpolation between noise and data, the Bayes-optimal velocity is governed to leading order at the noise end by a few class-conditional moments, and increasingly sample-specific structure matters toward the data end. StraightDP exploits this heterogeneity end to end. A small budget share releases whitened class-conditional moments once, to be distilled into the weights or injected at sampling time. The rest is spent by pre-declared DP-SGD toward the data end, beyond the moments' reach. At varepsilon=1 on MNIST, the released moments alone already attain 0.76 downstream accuracy with prototype-like samples and an FID of 237, and uniform DP-SGD attains 0.21. The pipeline built on the release reaches 0.81 accuracy at FID 56 in a public latent space. Constraining per-token stream norms of the multimodal backbone leaves the pretraining loss unchanged yet improves downstream accuracy in the extreme-noise pixel-space regime, and its accuracy effect becomes monotonically more favorable as privacy strengthens. The released moments also port to frozen SD3-medium, where sampling-time injection beats DP-LoRA training at a fraction of the budget.
Key facts. No citing papers in the map yet (data through 14 Oct 2026). Builds on 42 of its 63 references that are in the map. First posted on arXiv 31 Jul 2026.
Builds on
42 of 63 references.
References · 42
- 2017Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithmscited by 4,349
- 2024Scaling Rectified Flow Transformers for High-Resolution Image Synthesiscited by 4,181
- 2022Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flowcited by 3,700
- 2016Membership Inference Attacks against Machine Learning Modelscited by 2,281
- 2018Spectral Normalization for Generative Adversarial Networkscited by 1,636
- 2019Improved Precision and Recall Metric for Assessing Generative Modelscited by 944
- 2021Membership Inference Attacks From First Principlescited by 878
- 2018The Secret Sharer: Evaluating and Testing Unintended Memorization in Neural Networkscited by 864
- 2020Reliable Fidelity and Diversity Metrics for Generative Modelscited by 434
- 2018Deep Learning for Classical Japanese Literaturecited by 422
- 2018Assessing Generative Models via Precision and Recallcited by 406
- 2021Differentially Private Fine-tuning of Language Modelscited by 400
30 more
- 2016Concentrated Differential Privacy: Simplifications, Extensions, and Lower Boundscited by 346
- 2013The Composition Theorem for Differential Privacycited by 297
- 2022Rectified Flow: A Marginal Preserving Approach to Optimal Transportcited by 283
- 2018Privacy Amplification by Subsampling: Tight Analyses via Couplings and Divergencescited by 242
- 2018Subsampled R'enyi Differential Privacy and Analytical Moments Accountantcited by 236
- 2018Differentially Private Generative Adversarial Networkcited by 235
- 2020Differentially Private Learning Needs Better Features (or Much More Data)cited by 235
- 2022Unlocking High-Accuracy Differentially Private Image Classification through Scalecited by 224
- 2019Differentially Private Learning with Adaptive Clippingcited by 220
- 2020Auditing Differentially Private Machine Learning: How Private is Private SGD?cited by 218
- 2021Adversary Instantiation: Lower Bounds for Differentially Private Machine Learningcited by 199
- 2021Numerical Composition of Differential Privacycited by 180
- 2020The Lipschitz Constant of Self-Attentioncited by 174
- 2020GS-WGAN: A Gradient-Sanitized Approach for Learning Differentially Private Generatorscited by 117
- 2024Improving the Training of Rectified Flowscited by 108
- 2022Automatic Clipping: Differentially Private Deep Learning Made Easier and Strongercited by 88
- 2022Toward Training at ImageNet Scale with Differential Privacycited by 87
- 2023Differentially Private Diffusion Models Generate Useful Synthetic Imagescited by 86
- 2020DP-MERF: Differentially Private Mean Embeddings with Random Features for Practical Privacy-Preserving Data Generationcited by 78
- 2019AdaCliP: Adaptive Clipping for Private SGDcited by 77
- 2023Differentially Private Synthetic Data via Foundation Model APIs 1: Imagescited by 65
- 2021Don't Generate Me: Training Differentially Private Generative Models with Sinkhorn Divergencecited by 54
- 2019G-PATE: Scalable Differentially Private Data Generator via Private Aggregation of Teacher Discriminatorscited by 53
- 2022Connect the Dots: Tighter Discrete Approximations of Privacy Loss Distributionscited by 52
- 2022Large Scale Transfer Learning for Differentially Private Image Classificationcited by 41
- 2023Why Is Public Pretraining Necessary for Private Model Training?cited by 37
- 2021Dynamic Differential-Privacy Preserving SGDcited by 30
- 2021Hermite Polynomial Features for Private Data Generationcited by 21
- 2023Differentially Private Neural Tangent Kernels for Privacy-Preserving Data Generationcited by 15
- 2023DP-SGD Without Clipping: The Lipschitz Neural Network Waycited by 11
Built on by
Not cited yet.
Similar, not cited
Close in wording, no citation either way
- 2022Differentially Private Diffusion Models
- 2025From Easy to Hard: Building a Shortcut for Differentially Private Image Synthesis
- 2024DP-TLDM: Differentially Private Tabular Latent Diffusion Model
- 2026Protecting Sensitive Data in Image Synthesis via PAC-Private Adaptation for Diffusion Models
- 2024Differentially Private Fine-Tuning of Diffusion Models
- 2023DP-Forward: Fine-tuning and Inference on Language Models with Differential Privacy in Forward Pass
- 2022Directional Privacy for Deep Learning
- 2024Learning Differentially Private Diffusion Models via Stochastic Adversarial Distillation