aiqbench analysis
StraightDP addresses the utility cliff in differentially private generative models by leveraging the geometric properties of rectified flows to optimize budget allocation between global moments and sample-specific updates. This approach significantly improves downstream accuracy on MNIST compared to standard DP-SGD, even under strong privacy constraints ($\varepsilon=1$), while also demonstrating efficiency gains when applied to frozen SD3-medium backbones. The method effectively balances the release of class-conditional information with targeted training noise injection to maintain model fidelity in high-noise regimes.
StraightDP: Geometry-Aware Differential Privacy for Rectified-Flow Transformers
arXiv:2607.29100v1 Announce Type: new Abstract: 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.
Abstract: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.
Machine Learning (cs.LG); Cryptography and Security (cs.CR); Computer Vision and Pattern Recognition (cs.CV)
arXiv:2607.29100 [cs.LG]
(or arXiv:2607.29100v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2607.29100 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 301 words
Related coverage
Geometry-aware Latent Autoregressive Generative Model for PDEs in Complex Domains
arXiv · Sep 2, 2026
Luce: Relightable Gaussians for 3D Asset Generation
arXiv · Aug 27, 2026
VATO: A Vortex-Force-Aware Transformer Operator for Unsteady Separated Aerofoil Flows
arXiv · Sep 2, 2026
CityGuard: Graph-Aware Private Descriptors for Bias-Resilient Identity Search Across Urban Cameras
arXiv · Jul 23, 2026