arXiv:2510.22149v1 Announce Type: cross 
Abstract: Federated learning (FL) has emerged as a promising paradigm for decentralized model training, enabling multiple clients to collaboratively learn a shared model without exchanging their local data. However, the decentralized nature of FL also introduces vulnerabilities, as malicious clients can compromise or manipulate the training process. In this work, we introduce dictator clients, a novel, well-defined, and analytically tractable class of malicious participants capable of entirely erasing the contributions of all other clients from the server model, while preserving their own. We propose concrete attack strategies that empower such clients and systematically analyze their effects on the learning process. Furthermore, we explore complex scenarios involving multiple dictator clients, including cases where they collaborate, act independently, or form an alliance in order to ultimately betray one another. For each of these settings, we provide a theoretical analysis of their impact on the global model's convergence. Our theoretical algorithms and findings about the complex scenarios including multiple dictator clients are further supported by empirical evaluations on both computer vision and natural language processing benchmarks.

يتناول المقال مفهوم التعلم الفيدرالي (FL)، وهو نهج لامركزي لتدريب النماذج يسمح للعملاء بالتعاون دون تبادل بياناتهم. على الرغم من أن التعلم الفيدرالي يقدم مزايا كبيرة، إلا أنه يقدم أيضًا مخاطر، خاصة من العملاء الخبيثين الذين يمكنهم تعطيل عملية التدريب. تسلط مقدمة 'العملاء الديكتاتوريين' الضوء على نوع محدد من التهديدات ضمن هذا الإطار، مما يبرز الحاجة إلى استراتيجيات قوية لحماية نزاهة أنظمة التعلم الفيدرالي. فهم هذه الديناميكيات أمر بالغ الأهمية مع استمرار التعلم الفيدرالي في اكتساب الزخم في تطبيقات متنوعة.

El artículo discute el concepto de aprendizaje federado (FL), un enfoque descentralizado para el entrenamiento de modelos que permite a los clientes colaborar sin compartir sus datos. Aunque el FL ofrece ventajas significativas, también presenta riesgos, especialmente de clientes maliciosos que pueden interrumpir el proceso de entrenamiento. La introducción de 'clientes dictadores' destaca un tipo específico de amenaza dentro de este marco, enfatizando la necesidad de estrategias sólidas para salvaguardar la integridad de los sistemas de aprendizaje federado. Comprender estas dinámicas es crucial a medida que el FL continúa ganando terreno en diversas aplicaciones.

L'article aborde le concept d'apprentissage fédéré (FL), une approche décentralisée de l'entraînement de modèles qui permet aux clients de collaborer sans partager leurs données. Bien que le FL offre des avantages significatifs, il présente également des risques, notamment de la part de clients malveillants qui peuvent perturber le processus d'entraînement. L'introduction des 'clients dictateurs' met en lumière un type spécifique de menace au sein de ce cadre, soulignant la nécessité de stratégies robustes pour protéger l'intégrité des systèmes d'apprentissage fédéré. Comprendre ces dynamiques est crucial alors que le FL continue de gagner du terrain dans diverses applications.

The article discusses the concept of federated learning (FL), a decentralized approach to model training that allows clients to collaborate without sharing their data. While FL offers significant advantages, it also presents risks, particularly from malicious clients who can disrupt the training process. The introduction of 'dictator clients' highlights a specific type of threat within this framework, emphasizing the need for robust strategies to safeguard the integrity of federated learning systems. Understanding these dynamics is crucial as FL continues to gain traction in various applications.

Power to the Clients: Federated Learning in a Dictatorship Setting

arXiv:2511.18611v1 Announce Type: new 
Abstract: Split learning emerges as a promising paradigm for collaborative distributed model training, akin to federated learning, by partitioning neural networks between clients and a server without raw data exchange. However, sequential split learning suffers from poor scalability, while parallel variants like parallel split learning and split federated learning often incur high server resource overhead due to model duplication and aggregation, and generally exhibit reduced model performance and convergence owing to factors like client drift and lag. To address these limitations, we introduce CycleSL, a novel aggregation-free split learning framework that enhances scalability and performance and can be seamlessly integrated with existing methods. Inspired by alternating block coordinate descent, CycleSL treats server-side training as an independent higher-level machine learning task, resampling client-extracted features (smashed data) to mitigate heterogeneity and drift. It then performs cyclical updates, namely optimizing the server model first, followed by client updates using the updated server for gradient computation. We integrate CycleSL into previous algorithms and benchmark them on five publicly available datasets with non-iid data distribution and partial client attendance. Our empirical findings highlight the effectiveness of CycleSL in enhancing model performance. Our source code is available at https://gitlab.lrz.de/hctl/CycleSL.

تم تقديم CycleSL كإطار جديد للتعلم المنقسم القابل للتوسع، حيث يعالج القيود التي تواجه الأساليب الحالية من خلال القضاء على الحاجة إلى التجميع وتحسين الأداء. تتيح هذه الطريقة تعاونًا أفضل في تدريب النماذج الموزعة دون تبادل البيانات الخام، مما يحافظ على خصوصية البيانات.

CycleSL ha sido presentado como un nuevo marco para el aprendizaje dividido escalable, abordando las limitaciones de los métodos existentes al eliminar la necesidad de agregación y mejorar el rendimiento. Este enfoque permite una mejor colaboración en el entrenamiento de modelos distribuidos sin el intercambio de datos en bruto, manteniendo así la privacidad de los datos.

CycleSL a été introduit comme un nouveau cadre pour l'apprentissage fractionné évolutif, abordant les limitations des méthodes existantes en éliminant le besoin d'agrégation et en améliorant les performances. Cette approche permet une meilleure collaboration dans l'entraînement de modèles distribués sans échange de données brutes, préservant ainsi la confidentialité des données.

CycleSL has been introduced as a new framework for scalable split learning, addressing the limitations of existing methods by eliminating the need for aggregation and enhancing performance. This approach allows for improved collaboration in distributed model training without the exchange of raw data, thereby maintaining data privacy.

CycleSL: Server-Client Cyclical Update Driven Scalable Split Learning

arXiv:2511.19248v1 Announce Type: cross 
Abstract: Test-time personalization in federated learning enables models at clients to adjust online to local domain shifts, enhancing robustness and personalization in deployment. Yet, existing federated learning work largely overlooks the security risks that arise when local adaptation occurs at test time. Heterogeneous domain arrivals, diverse adaptation algorithms, and limited cross-client visibility create vulnerabilities where compromised participants can craft poisoned inputs and submit adversarial updates that undermine both global and per-client performance. To address this threat, we introduce FedPoisonTTP, a realistic grey-box attack framework that explores test-time data poisoning in the federated adaptation setting. FedPoisonTTP distills a surrogate model from adversarial queries, synthesizes in-distribution poisons using feature-consistency, and optimizes attack objectives to generate high-entropy or class-confident poisons that evade common adaptation filters. These poisons are injected during local adaptation and spread through collaborative updates, leading to broad degradation. Extensive experiments on corrupted vision benchmarks show that compromised participants can substantially diminish overall test-time performance.

تم تقديم إطار جديد يسمى FedPoisonTTP لمعالجة الثغرات الأمنية في التعلم الفيدرالي خلال تخصيص وقت الاختبار. يبرز هذا الإطار كيف يمكن للمشاركين المفسدين استغلال التكيفات المحلية لتقديم مدخلات مسمومة، مما قد يؤدي إلى تدهور الأداء العالمي والفردي للنماذج. يقوم الإطار بتوليف تحديثات عدائية لإنشاء سموم عالية الإنتروبيا أو موثوقة في الفئة، مما يشكل مخاطر كبيرة على نزاهة أنظمة التعلم الفيدرالي.

Se ha introducido un nuevo marco llamado FedPoisonTTP para abordar las vulnerabilidades de seguridad en el aprendizaje federado durante la personalización en el momento de la prueba. Este marco destaca cómo los participantes comprometidos pueden explotar las adaptaciones locales para enviar entradas envenenadas, lo que puede degradar tanto el rendimiento global como el individual de los modelos. El marco sintetiza actualizaciones adversariales para crear venenos de alta entropía o confiados en clase, planteando riesgos significativos para la integridad de los sistemas de aprendizaje federado.

Un nouveau cadre appelé FedPoisonTTP a été introduit pour aborder les vulnérabilités de sécurité dans l'apprentissage fédéré lors de la personnalisation au moment du test. Ce cadre met en évidence comment des participants compromis peuvent exploiter les adaptations locales pour soumettre des entrées empoisonnées, ce qui peut dégrader à la fois la performance globale et individuelle des modèles. Le cadre synthétise des mises à jour adversariales pour créer des poisons à haute entropie ou confiants en classe, posant des risques significatifs pour l'intégrité des systèmes d'apprentissage fédéré.

A new framework called FedPoisonTTP has been introduced to address security vulnerabilities in federated learning during test-time personalization. This framework highlights how compromised participants can exploit local adaptations to submit poisoned inputs, which can degrade both global and individual model performance. The framework synthesizes adversarial updates to create high-entropy or class-confident poisons, posing significant risks to the integrity of federated learning systems.

Power to the Clients: Federated Learning in a Dictatorship Setting

Was this article worth reading? Share it

Agentcloud

LucidQuery AI

DeployMCP