arXiv:2603.13421v2 Announce Type: replace 
Abstract: Generative models based on the Flow Matching objective, particularly Rectified Flow, have emerged as a dominant paradigm for efficient, high-fidelity image synthesis. However, while existing research heavily prioritizes generation quality and architectural scaling, the underlying dynamics of how RF models memorize training data remain largely underexplored. In this paper, we systematically investigate the memorization behaviors of RF through the test statistics of Membership Inference Attacks (MIA). We progressively formulate three test statistics, culminating in a complexity-calibrated metric that successfully decouples intrinsic image spatial complexity from genuine memorization signals. This calibration yields a significant performance surge -- boosting attack AUC by up to 15% and the privacy-critical TPR@1%FPR metric by up to 45% -- establishing the first non-trivial MIA specifically tailored for RF. Leveraging these refined metrics, we uncover a distinct temporal pattern: under standard uniform temporal training, a model's susceptibility to MIA strictly peaks at the integration midpoint, a phenomenon we justify via the network's forced deviation from linear approximations. Finally, we demonstrate that substituting uniform timestep sampling with a Symmetric Exponential (U-shaped) distribution effectively minimizes exposure to vulnerable intermediate timesteps. Extensive evaluations across three datasets confirm that this temporal regularization suppresses memorization while preserving generative fidelity.

تركزت الأبحاث الأخيرة على ديناميكيات نماذج التدفق المستقيم (RF)، وخاصة سلوكيات الذاكرة الخاصة بها، من خلال عدسة هجمات استدلال العضوية (MIA). تقدم الدراسة مقياسًا مُعَدَّلًا حسب التعقيد يميز بين التعقيد المكاني الداخلي للصورة وإشارات الذاكرة الحقيقية، مما يؤدي إلى زيادة ملحوظة في أداء مقاييس الهجوم.

Investigaciones recientes se han centrado en la dinámica de los modelos de Flujo Rectificado (RF), especialmente en sus comportamientos de memorización, a través de las Ataques de Inferencia de Membresía (MIA). El estudio introduce una métrica calibrada por complejidad que distingue entre la complejidad intrínseca de la imagen y las señales de memorización reales, lo que resulta en un aumento notable en las métricas de rendimiento de ataque.

Des recherches récentes se sont concentrées sur la dynamique des modèles de Flux Rectifié (RF), en particulier leurs comportements de mémorisation, à travers le prisme des Attaques d'Inférence de Membre (MIA). L'étude introduit une métrique calibrée par la complexité qui distingue entre la complexité intrinsèque de l'image et les signaux de mémorisation réels, entraînant une augmentation notable des performances des métriques d'attaque.

Recent research has focused on the dynamics of Rectified Flow (RF) models, particularly their memorization behaviors, through the lens of Membership Inference Attacks (MIA). The study introduces a complexity-calibrated metric that distinguishes between intrinsic image complexity and actual memorization signals, resulting in a notable increase in attack performance metrics.