arXiv:2504.05822v2 Announce Type: replace 
Abstract: The right to be forgotten is a fundamental principle of privacy-preserving regulations and extends to Machine Learning (ML) paradigms such as Federated Learning (FL). While FL enhances privacy by enabling collaborative model training without sharing private data, trained models still retain the influence of training data. Federated Unlearning (FU) methods recently proposed often rely on impractical assumptions for real-world FL deployments, such as storing client update histories or requiring access to a publicly available dataset. To address these constraints, this paper introduces a novel method that leverages negated Pseudo-gradients Updates for Federated Unlearning (PUF). Our approach only uses standard client model updates, which are employed during regular FL rounds, and interprets them as pseudo-gradients. When a client needs to be forgotten, we apply the negation of their pseudo-gradients, appropriately scaled, to the global model. Unlike state-of-the-art mechanisms, PUF seamlessly integrates with FL workflows, incurs no additional computational and communication overhead beyond standard FL rounds, and supports concurrent unlearning requests. We extensively evaluated the proposed method on two well-known benchmark image classification datasets (CIFAR-10 and CIFAR-100) and a real-world medical imaging dataset for segmentation (ProstateMRI), using three different neural architectures: two residual networks and a vision transformer. The experimental results across various settings demonstrate that PUF achieves state-of-the-art forgetting effectiveness and recovery time, without relying on any additional assumptions.

تناقش ورقة حديثة التنفيذ العملي لأسلوب النسيان الفيدرالي (FU)، وهو تقدم حاسم في التعلم الآلي الذي يحترم الخصوصية. تتناول هذه الطريقة الحق في النسيان من خلال السماح للنماذج بنسيان بيانات معينة دون التأثير على أدائها العام. مع تزايد أهمية خصوصية البيانات، فإن القدرة على دمج FU بسلاسة في أنظمة التعلم الفيدرالي (FL) يمكن أن تحدث ثورة في كيفية تعامل المؤسسات مع المعلومات الحساسة، مما يمثل خطوة مهمة نحو ممارسات الذكاء الاصطناعي الأخلاقية.

Un artículo reciente discute la implementación práctica del Olvido Federado (FU), un avance crucial en el aprendizaje automático que respeta la privacidad. Este método aborda el derecho al olvido al permitir que los modelos olviden datos específicos sin comprometer su rendimiento general. A medida que la privacidad de los datos se vuelve cada vez más importante, la capacidad de integrar FU de manera fluida en los sistemas de Aprendizaje Federado (FL) podría revolucionar la forma en que las organizaciones manejan información sensible, lo que representa un paso significativo hacia prácticas éticas en la IA.

Un article récent aborde la mise en œuvre pratique de l'oubli fédéré (FU), une avancée cruciale dans l'apprentissage automatique respectueux de la vie privée. Cette méthode traite du droit à l'oubli en permettant aux modèles d'oublier des données spécifiques sans compromettre leur performance globale. Alors que la confidentialité des données devient de plus en plus importante, la capacité d'intégrer de manière transparente le FU dans les systèmes d'apprentissage fédéré (FL) pourrait révolutionner la manière dont les organisations gèrent les informations sensibles, représentant ainsi une avancée significative dans les pratiques éthiques de l'IA.

A recent paper discusses the practical implementation of Federated Unlearning (FU), a crucial advancement in privacy-preserving machine learning. This method addresses the right to be forgotten by allowing models to forget specific data without compromising overall performance. As data privacy becomes increasingly important, the ability to seamlessly integrate FU into Federated Learning (FL) systems could revolutionize how organizations handle sensitive information, making it a significant step forward in ethical AI practices.

Federated Unlearning Made Practical: Seamless Integration via Negated Pseudo-Gradients

arXiv:2507.06819v3 Announce Type: replace 
Abstract: Prototype models are an important method for explainable artificial intelligence (XAI) and interpretable machine learning. In this paper, we perform an in-depth analysis of a set of prominent prototype models including ProtoPNet, ProtoPool and PIPNet. For their assessment, we apply a comprehensive set of metrics. In addition to applying standard metrics from literature, we propose several new metrics to further complement the analysis of model interpretability. In our experimentation, we apply the set of prototype models on a diverse set of datasets including fine-grained classification, Non-IID settings and multi-label classification to further contrast the performance. Furthermore, we also provide our code as an open-source library (https://github.com/uos-sis/quanproto), which facilitates simple application of the metrics itself, as well as extensibility -- providing the option for easily adding new metrics and models.

تم إجراء تقييم شامل لشبكات الأعصاب النموذجية، مع التركيز على نماذج مثل ProtoPNet وProtoPool وPIPNet. تطبق الدراسة مجموعة متنوعة من المقاييس، بما في ذلك مقاييس جديدة اقترحها المؤلفون، لتقييم قابلية تفسير النموذج عبر مجموعات بيانات متنوعة، بما في ذلك مهام التصنيف الدقيق ومتعدد التسميات. الكود الخاص بهذه التقييمات متاح كمكتبة مفتوحة المصدر على GitHub.

Se ha realizado una evaluación integral de redes neuronales prototipo, centrándose en modelos como ProtoPNet, ProtoPool y PIPNet. El estudio aplica una variedad de métricas, incluidas algunas nuevas propuestas por los autores, para evaluar la interpretabilidad del modelo en diversos conjuntos de datos, incluidas tareas de clasificación fina y multi-etiqueta. El código para estas evaluaciones está disponible como una biblioteca de código abierto en GitHub.

Une évaluation complète des réseaux de neurones prototypes a été réalisée, en se concentrant sur des modèles tels que ProtoPNet, ProtoPool et PIPNet. L'étude applique une variété de métriques, y compris de nouvelles proposées par les auteurs, pour évaluer l'interprétabilité des modèles sur divers ensembles de données, y compris des tâches de classification fine et multi-label. Le code pour ces évaluations est disponible en tant que bibliothèque open-source sur GitHub.

A comprehensive evaluation of prototype neural networks has been conducted, focusing on models such as ProtoPNet, ProtoPool, and PIPNet. The study applies a variety of metrics, including new ones proposed by the authors, to assess model interpretability across diverse datasets, including fine-grained and multi-label classification tasks. The code for these evaluations is available as an open-source library on GitHub.

Comprehensive Evaluation of Prototype Neural Networks

arXiv:2511.16792v1 Announce Type: cross 
Abstract: Membership inference attacks (MIAs) against machine learning (ML) models aim to determine whether a given data point was part of the model training data. These attacks may pose significant privacy risks to individuals whose sensitive data were used for training, which motivates the use of defenses such as differential privacy, often at the cost of high accuracy losses. MIAs exploit the differences in the behavior of a model when making predictions on samples it has seen during training (members) versus those it has not seen (non-members). Several studies have pointed out that model overfitting is the major factor contributing to these differences in behavior and, consequently, to the success of MIAs. However, the literature also shows that even non-overfitted ML models can leak information about a small subset of their training data. In this paper, we investigate the root causes of membership inference vulnerabilities beyond traditional overfitting concerns and suggest targeted defenses. We empirically analyze the characteristics of the training data samples vulnerable to MIAs in models that are not overfitted (and hence able to generalize). Our findings reveal that these samples are often outliers within their classes (e.g., noisy or hard to classify). We then propose potential defensive strategies to protect these vulnerable samples and enhance the privacy-preserving capabilities of ML models. Our code is available at https://github.com/najeebjebreel/mia_analysis.

تثير هجمات استنتاج العضوية (MIA) ضد نماذج التعلم الآلي مخاوف كبيرة بشأن الخصوصية، حيث يمكنها تحديد ما إذا كانت نقاط بيانات معينة جزءًا من بيانات التدريب. تستكشف هذه الورقة نقاط الضعف في MIA التي تستمر حتى في النماذج غير المفرطة التكيف، مما يبرز الحاجة إلى تحسين الدفاعات بما يتجاوز الأساليب التقليدية مثل الخصوصية التفاضلية.

Los ataques de inferencia de pertenencia (MIA) contra modelos de aprendizaje automático han suscitado preocupaciones significativas sobre la privacidad, ya que pueden determinar si puntos de datos específicos formaron parte de los conjuntos de datos de entrenamiento. Este artículo investiga las vulnerabilidades a los MIA que persisten incluso en modelos no sobreajustados, destacando la necesidad de defensas mejoradas más allá de los métodos tradicionales como la privacidad diferencial.

Les attaques par inférence d'appartenance (MIA) contre les modèles d'apprentissage automatique soulèvent d'importantes préoccupations en matière de confidentialité, car elles peuvent déterminer si des points de données spécifiques ont été inclus dans les ensembles de données d'entraînement. Cet article explore les vulnérabilités aux MIA qui persistent même dans les modèles non surajustés, soulignant la nécessité d'améliorer les défenses au-delà des méthodes traditionnelles comme la confidentialité différentielle.

Membership inference attacks (MIAs) against machine learning models have raised significant privacy concerns, as they can determine if specific data points were included in training datasets. This paper explores vulnerabilities to MIAs that persist even in non-overfitted models, highlighting the need for improved defenses beyond traditional methods like differential privacy.

Federated Unlearning Made Practical: Seamless Integration via Negated Pseudo-Gradients

Was this article worth reading? Share it

Teilord

Lenso.ai

Adaptive Privacy Policy Generator