Where Flow Matching Leaks: Characterising the Membership Signal Along the Interpolation Path
Published in ICML 20226, 2026
Flow Matching leaks information about their training data, and they leak it most exactly where they know the least. We characterise this membership signal along the interpolation path, predict its peak in closed form, and show it stays invisible to standard training metrics
Recommended citation: sesmat2026where, title={Where Flow Matching Leaks: Characterising Membership Signals Along the Interpolation Path}, author={Thomas Sesmat and Gabriel Meseguer-Brocal and Geoffroy Peeters}, booktitle={Forty-third International Conference on Machine Learning}, year={2026}, url={https://openreview.net/forum?id=Ty5X41WbJw}