Back to search

National Technical University of Athens (NTUA)

A multi-agent distributed learning framework for the optimization of privacy and resilience in federated systems

Abstract

dc:description

Federated learning has emerged as a foundational paradigm for collaborative model training across distributed agents without explicit data sharing. However, as federated systems scale to real-world deployments, they face fundamental challenges arising from data heterogeneity, infrastructure constraints, privacy requirements, and adversarial conditions. Existing approaches typically address these challenges in isolation, failing to account for their complex interdependencies. This thesis develops a multi-agent distributed learning framework that jointly optimizes privacy and resilience in federated systems by explicitly modeling agent interactions, structural dependencies, and system-level dynamics. The central argument of this thesis is that federated learning must be understood not as a static optimization protocol, but as an adaptive, structured, multi-agent process in which privacy, resilience, efficiency, and coordination are deeply intertwined. By treating participating agents as heterogeneous entities with distinct roles, influence, and constraints, and by leveraging graph-based representations to capture their interactions, the proposed framework enables principled reasoning about privacy-utility trade-offs, robust aggregation under adversarial conditions, and scalable coordination in dynamic environments. To realize this vision, the thesis develops a progression of interrelated contributions: from foundational coordination mechanisms that enable structure-aware agent organization, through adaptive privacy protocols and secure collaboration primitives, to efficiency-enhancing optimization techniques and their integration into a unified system architecture. Each contribution addresses a specific dimension of the federated learning challenge while maintaining coherence with the overarching multi-agent framework. The following overview summarizes the content and contributionsof each chapter. Chapter 2 establishes the theoretical and technical foundations underlying the thesis. It reviews federated learning fundamentals, including core algorithms such as FedAvg and FedProx, statistical heterogeneity challenges, and privacy-preserving mechanisms including differential privacy, secure multi-party computation, and homomorphic encryption. The chapter further examines security threats in federated systems, graph neural networks for decentralized environments, and cybersecurity challenges in large-scale Internet of Things ecosystems. This comprehensive review situates the thesis contributions within the broader landscape of distributed machine learning and privacy-preserving computation. Chapter 3 introduces structure-aware clustering mechanisms that organize heterogeneous agents based on spatial, resource, and interaction characteristics. Motivated by smart-city scenarios involving mobile digital twin boxes equipped with 360 degree video streaming capabilities, the chapter develops unsupervised learning approaches combining density-based clustering, multi-view spectral clustering, and k-nearest neighbor allocation to maximize situational awareness under bandwidth constraints. The proposed framework jointly optimizes area coverage and video quality allocation, demonstrating significant improvements over baseline approaches across varying vehicle densities. These clustering mechanisms serve as the foundational coordination layer for the multi-agent federated learning framework developed in subsequent chapters. Chapter 4 addresses privacy preservation in federated graph learning through two complementary contributions. First, it introduces an adaptive differential privacy mechanism that dynamically adjusts privacy budgets based on local graph structure, including edge density and node degree distributions. Unlike conventional fixed-noise approaches, this graph-aware calibration achieves improved privacy-utility trade-offs by reducing excessive noise for well-connected graphs while maintaining strict guarantees for sparse structures. Experiments on medical imaging datasets demonstrate faster convergence and higher accuracy compared to fixed differential privacy baselines. Second, the chapter presents a threshold-based multiparty private set intersection (T-MPSI) framework that enables flexible secure collaboration by including elements appearing in at least a threshold number of participant datasets, rather than requiring exact matching. The framework employs federated graph neural networks for adaptive threshold selection and integrates homomorphic encryption for protecting intermediate computations. Chapter 5 develops graph neural network methods for optimizing client participation under differential privacy constraints. Recognizing that client selection directly impacts both learning efficiency and privacy exposure, the chapter formulates client selection as a structure-aware optimization problem that accounts for data heterogeneity, resource constraints, and privacy budgets. The proposed approach leverages graph attention mechanisms to learn client embeddings that capture both local data characteristics and system-level relationships, enabling adaptive participation strategies that improve convergence while maintaining formal privacy guarantees. Extensive experiments demonstrate the effectiveness of the approach in balancing privacy, utility, and computational efficiency across heterogeneous federated environments. Chapter 6 explores quantum-inspired optimization techniques for improving the efficiency and scalability of federated learning under privacy constraints. The chapter introduces QuanGAT, a hybrid framework integrating quantum neural networks, graph attention networks, and federated learning for privacy-preserving analysis of biological graph data. QuanGAT employs parameterized quantum circuits with angle encoding and quantum noise channels to model biological variability, while incorporating Laplace noise injection in the attention mechanism to enforce differential privacy. Evaluated on protein–protein interaction networks for DNA mutation prediction, QuanGAT consistently outperforms baseline graph neural networks in federated settings maintaining robustness to noise and client heterogeneity. Chapter 7 integrates the individual contributions into a cohesive, verifiable, and Byzantine-resilient federated learning architecture. The unified framework operates under an explicit adversarial model that accounts for Byzantine clients, honest-but-curious servers, and network-level threats. Resilience is achieved through layered mitigation strategies spanning adaptive client selection, structure-aware aggregation, trust-weighted participation, and anomaly detection in embedding spaces. The architecture combines adaptive differential privacy, threshold-based secure operations, and graph-based coordination to simultaneously satisfy privacy, robustness, and scalability requirements. Verifiable aggregation mechanisms enable post-hoc accountability without sacrificing privacy by default, ensuring that aggregation correctness can be cryptographically certified while individual contributions remain protected. The final chapter summarizes the contributions of the thesis, discusses limitations, and outlines directions for future research, including extensions to cross-silo federated learning, integration with emerging privacy-enhancing technologies, and deployment considerations for real-world distributed systems.The contributions of this thesis advance the state of the art in federated learning by demonstrating that privacy, resilience, and efficiency can be jointly optimized within a structure-aware, multi-agent framework. Key theoretical findings include: (i) structure-dependent privacy amplification, where noise calibration exploits topological redundancy to improve privacy-utility trade-offs while maintaining formal ε,δ -differential privacy guarantees; (ii) composability of privacy-preserving and Byzantine-resilient mechanisms through layered architectural design, resolving the apparent tension between hiding individual contributions and detecting adversarial behavior; and (iii) variance reduction in aggregated updates through importance-weighted client participation, accelerating convergence under non-IID data distributions. The proposed mechanisms, such as adaptive graph-aware privacy calibration, threshold-based secure collaboration, quantum-enhanced graph learning, and verifiable Byzantine-resilient aggregation, provide principled solutions to challenges that have traditionally been addressed in isolation. Together, they form a cohesive framework for designing federated systems that remain trustworthy, robust, and scalable under realistic deployment conditions.

Degree

thesis:*
Grantor dc:publisher
National Technical University of Athens (NTUA)
Year dc:date
2026

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Messinis, Sotirios
  • Μεσσήνης, Σωτήριος

Subjects

dc:subject × 22

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/62224

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Messinis, Sotirios; Μεσσήνης, Σωτήριος. A multi-agent distributed learning framework for the optimization of privacy and resilience in federated systems. National Technical University of Athens (NTUA), 2026. http://hdl.handle.net/10442/hedi/62224