Where Rectified Flows Leak: Characterising Membership Signals Along the Interpolation Path

Hugging Face Daily Papers Papers

Summary

This paper studies membership inference attacks on Rectified Flows by analyzing the interpolation path, revealing a bell-shaped gap between train and test data reconstruction that accumulates during training.

Understanding what generative models retain from training data remains challenging, with implications for copyright and privacy. Beyond verbatim reproduction, models can encode subtler traces of their training data that never surface in their outputs yet remain exploitable. We study this regime for Rectified Flows, which are increasingly used in deployed generative systems. We analyse the interpolation path X_λ= (1-λ)X_0 + λX_1 that defines the Rectified Flow training. We show that a gap exists between the reconstruction of train and test data that follows a bell-shaped curve over λ, wich accumulates during training, while the validation metrics remain stable. The signal has a maximum whose location we derive in closed form under Gaussian assumptions. We validate these predictions on both audio and images and show that the bell-shaped structure is universal, while the peak prediction holds when our assumptions are satisfied. As a proof of concept, we exploit this specific λ-resolved structure to perform a Membership Inference Attack, distinguishing members of the training set from non-members.
Original Article
View Cached Full Text

Cached at: 06/10/26, 12:13 AM

Paper page - Where Rectified Flows Leak: Characterising Membership Signals Along the Interpolation Path

Source: https://huggingface.co/papers/2606.07271

Abstract

Rectified Flows retain subtle training data traces that accumulate during training and can be exploited for membership inference attacks.

Understanding what generative models retain from training data remains challenging, with implications for copyright and privacy. Beyond verbatim reproduction, models can encode subtler traces of their training data that never surface in their outputs yet remain exploitable. We study this regime forRectified Flows, which are increasingly used in deployed generative systems. We analyse theinterpolation pathX_λ= (1-λ)X_0 + λX_1 that defines the Rectified Flow training. We show that a gap exists between the reconstruction of train and test data that follows abell-shaped curveover λ, wich accumulates during training, while the validation metrics remain stable. The signal has a maximum whose location we derive in closed form underGaussian assumptions. We validate these predictions on both audio and images and show that the bell-shaped structure is universal, while the peak prediction holds when our assumptions are satisfied. As a proof of concept, we exploit this specific λ-resolved structure to perform aMembership Inference Attack, distinguishing members of the training set from non-members.

View arXiv pageView PDFGitHub0Add to collection

Get this paper in your agent:

hf papers read 2606\.07271

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2606.07271 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2606.07271 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2606.07271 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

MeshFlow: Mesh Generation with Equivariant Flow Matching

Hugging Face Daily Papers

MeshFlow introduces an equivariant optimal-transport flow matching model for direct triangle mesh generation, achieving state-of-the-art quality while providing approximately 18x inference speedup over autoregressive methods.

Probabilistic Inversion with Flow Matching

arXiv cs.LG

This paper applies Flow Matching, a generative AI technique, to probabilistic inversion for seismic full-waveform inversion, demonstrating its effectiveness on synthetic datasets.

Membership Inference Attacks on Discrete Diffusion Language Models

arXiv cs.LG

This paper studies membership inference attacks (MIA) on fine-tuned masked diffusion language models (MDLMs). It proposes a white-box attack using a 46-dimensional feature vector from the model's reconstruction loss at varying masking ratios, achieving high AUC scores and showing MDLMs are more vulnerable than previously thought.