arXiv:2507.18725v2 Announce Type: replace 
Abstract: Machine unlearning aims to efficiently eliminate the memory about deleted data from trained models and address the right to be forgotten. Despite the success of existing unlearning algorithms, unlearning in sparse models has not yet been well studied. In this paper, we empirically find that the deleted data has an impact on the pruned topology in a sparse model. Motivated by the observation and the right to be forgotten, we define a new terminology ``un-pruning" to eliminate the impact of deleted data on model pruning. Then we propose an un-pruning algorithm to approximate the pruned topology driven by retained data. We remark that any existing unlearning algorithm can be integrated with the proposed un-pruning workflow and the error of un-pruning is upper-bounded in theory. Also, our un-pruning algorithm can be applied to both structured sparse models and unstructured sparse models. In the experiment, we further find that Membership Inference Attack (MIA) accuracy is unreliable for assessing whether a model has forgotten deleted data, as a small change in the amount of deleted data can produce arbitrary MIA results. Accordingly, we devise new performance metrics for sparse models to evaluate the success of un-pruning. Lastly, we conduct extensive experiments to verify the efficacy of un-pruning with various pruning methods and unlearning algorithms. Our code is released at https://github.com/NKUShaw/SparseModels .

تقدم دراسة حديثة مفهوم 'إزالة التقليم' في التعلم الآلي، مع التركيز على كيفية تأثير البيانات المحذوفة على طوبولوجيا النماذج المتناثرة. تهدف هذه الطريقة الجديدة إلى إزالة الذاكرة المتعلقة بالبيانات المحذوفة من النماذج المدربة بشكل فعال، مع معالجة الحق في النسيان. يمكن دمج خوارزمية إزالة التقليم المقترحة مع طرق التعلم الموجودة، وهي قابلة للتطبيق على كل من النماذج المتناثرة الهيكلية وغير الهيكلية.

Un estudio reciente presenta el concepto de 'despodar' en el aprendizaje automático, centrándose en cómo los datos eliminados influyen en la topología de los modelos dispersos. Este nuevo enfoque busca eliminar de manera efectiva la memoria de los datos eliminados de los modelos entrenados, abordando el derecho al olvido. El algoritmo de despodado propuesto puede integrarse con métodos de aprendizaje existentes y es aplicable tanto a modelos dispersos estructurados como no estructurados.

Une étude récente introduit le concept de 'dé-pruning' dans l'apprentissage machine, en se concentrant sur la manière dont les données supprimées influencent la topologie des modèles épars. Cette nouvelle approche vise à éliminer efficacement la mémoire des données supprimées des modèles entraînés, en abordant le droit à l'oubli. L'algorithme de dé-pruning proposé peut être intégré aux méthodes d'apprentissage existantes et est applicable aux modèles épars structurés et non structurés.

A recent study introduces the concept of 'un-pruning' in machine unlearning, focusing on how deleted data influences the topology of sparse models. This new approach aims to effectively eliminate the memory of deleted data from trained models, addressing the right to be forgotten. The proposed un-pruning algorithm can be integrated with existing unlearning methods and is applicable to both structured and unstructured sparse models.

The Right to be Forgotten in Pruning: Unveil Machine Unlearning on Sparse Models

arXiv:2512.08443v1 Announce Type: new 
Abstract: Machine unlearning (MU) seeks to remove the influence of specified data from a trained model in response to privacy requests or data poisoning. While certified unlearning has been analyzed in centralized and server-orchestrated federated settings (via guarantees analogous to differential privacy, DP), the decentralized setting -- where peers communicate without a coordinator remains underexplored. We study certified unlearning in decentralized networks with fixed topologies and propose RR-DU, a random-walk procedure that performs one projected gradient ascent step on the forget set at the unlearning client and a geometrically distributed number of projected descent steps on the retained data elsewhere, combined with subsampled Gaussian noise and projection onto a trust region around the original model. We provide (i) convergence guarantees in the convex case and stationarity guarantees in the nonconvex case, (ii) $(\varepsilon,\delta)$ network-unlearning certificates on client views via subsampled Gaussian $R\'enyi$ DP (RDP) with segment-level subsampling, and (iii) deletion-capacity bounds that scale with the forget-to-local data ratio and quantify the effect of decentralization (network mixing and randomized subsampling) on the privacy--utility trade-off. Empirically, on image benchmarks (MNIST, CIFAR-10), RR-DU matches a given $(\varepsilon,\delta)$ while achieving higher test accuracy than decentralized DP baselines and reducing forget accuracy to random guessing ($\approx 10\%$).

قدمت دراسة حديثة طريقة للتعلم غير المراقب المعتمد بالكامل في بيئات غير مركزية، تركز على إزالة تأثيرات بيانات محددة من النماذج المدربة دون الحاجة إلى منسق مركزي. تُعرف هذه الطريقة باسم RR-DU، وتستخدم إجراءً عشوائيًا لتحسين الخصوصية وتقليل مخاطر تسميم البيانات، مع توفير ضمانات التقارب في السيناريوهات المحدبة وضمانات الاستقرار في الحالات غير المحدبة.

Un estudio reciente ha presentado un método para el desaprendizaje certificado totalmente descentralizado en el aprendizaje automático, centrado en la eliminación de influencias de datos específicos de modelos entrenados sin un coordinador central. Este enfoque, denominado RR-DU, emplea un procedimiento de caminata aleatoria para mejorar la privacidad y mitigar los riesgos de envenenamiento de datos, proporcionando garantías de convergencia en escenarios convexos y de estacionariedad en casos no convexos.

Une étude récente a introduit une méthode pour un apprentissage non supervisé certifié entièrement décentralisé dans l'apprentissage automatique, axée sur la suppression des influences de données spécifiques des modèles entraînés sans un coordinateur central. Cette approche, appelée RR-DU, utilise une procédure de marche aléatoire pour améliorer la confidentialité et atténuer les risques de contamination des données, fournissant des garanties de convergence dans des scénarios convexes et de stationnarité dans des cas non convexes.

A recent study has introduced a method for fully decentralized certified unlearning in machine learning, focusing on the removal of specific data influences from trained models without a central coordinator. This approach, termed RR-DU, employs a random-walk procedure to enhance privacy and mitigate data poisoning risks, providing convergence guarantees in convex scenarios and stationarity in nonconvex cases.

The Right to be Forgotten in Pruning: Unveil Machine Unlearning on Sparse Models

Was this article worth reading? Share it

LucidQuery AI

AskTuring

Hypertune