arXiv:2507.16880v3 Announce Type: replace-cross 
Abstract: Text-to-image diffusion models (DMs) have achieved remarkable success in image generation. However, concerns about data privacy and intellectual property remain due to their potential to inadvertently memorize and replicate training data. Recent mitigation efforts have focused on identifying and pruning weights responsible for triggering verbatim training data replication, based on the assumption that memorization can be localized. We challenge this assumption and demonstrate that, even after such pruning, small perturbations to the text embeddings of previously mitigated prompts can re-trigger data replication, revealing the fragility of such methods. Our further analysis then provides multiple indications that memorization is indeed \textit{not} inherently local: (1) replication triggers for memorized images are distributed throughout text embedding space; (2) embeddings yielding the same replicated image produce divergent model activations; and (3) different pruning methods identify inconsistent sets of memorization-related weights for the same image. Finally, we show that bypassing the locality assumption enables more robust mitigation through adversarial fine-tuning. These findings provide new insights into the fundamental nature of memorization in text-to-image DMs and inform the future development of more reliable mitigation methods against DM memorization.

دراسة حديثة بعنوان 'Finding DoRI: Discovery of Retained Images in Diffusion Models' تسلط الضوء على المخاوف المتعلقة بخصوصية البيانات وحقوق الملكية الفكرية في نماذج الانتشار من النص إلى الصورة، حيث تكشف أنه حتى بعد محاولات التخفيف من حفظ بيانات التدريب، يمكن أن تؤدي التغييرات الصغيرة في النصوص إلى إعادة تفعيل تكرار البيانات.

Un estudio reciente titulado 'Finding DoRI: Discovery of Retained Images in Diffusion Models' destaca preocupaciones sobre la privacidad de los datos y la propiedad intelectual en los modelos de difusión de texto a imagen, revelando que incluso después de los intentos de mitigar la memorización de los datos de entrenamiento, pequeños cambios en los mensajes de texto pueden volver a activar la replicación de datos.

Une étude récente intitulée 'Finding DoRI: Discovery of Retained Images in Diffusion Models' met en lumière des préoccupations concernant la confidentialité des données et la propriété intellectuelle dans les modèles de diffusion d'images à partir de texte, révélant qu'après des tentatives de mitigation de la mémorisation des données d'entraînement, de petits changements dans les invites textuelles peuvent encore déclencher la réplication des données.

A recent study titled 'Finding DoRI: Discovery of Retained Images in Diffusion Models' highlights concerns regarding data privacy and intellectual property in text-to-image diffusion models, revealing that even after attempts to mitigate memorization of training data, small changes to text prompts can still trigger data replication.