Back to search

Democritus University of Thrace (DUTH)

Optimizing responsible decentralized machine learning methods for AI applications

Abstract

dc:description

The exponential growth of artificial intelligence technologies has led to their widespread adoption across a diverse range of domains, including healthcare, telecommunications, energy, and personalized digital services. As AI systems become increasingly integrated into critical infrastructures and real-world decision-making processes, important ethical and technical concerns have emerged. These include risks to data privacy, lack of transparency in model decisions, fairness in algorithmic outcomes, and the rising environmental footprint of large-scale machine learning models. Addressing these concerns requires a shift from purely performance-driven AI to a more Responsible AI paradigm, an approach that prioritizes ethical, sustainable, and human-centric principles in the design and deployment of intelligent systems. This thesis approaches Responsible AI through three core pillars: privacy, explainability, and sustainability. It investigates how modern machine learning pipelines can be designed to meet these principles in practice, by developing methods that are not only accurate but also ethically and socially aligned. A central focus of this thesis is the application of Responsible AI principles in decentralized settings, where data privacy, system heterogeneity, and lack of centralized control present unique challenges. In this context, Federated Learning is investigated as a primary technique to apply these principles across diverse domains. Federated Learning allows multiple data-holding parties (for example, edge devices or institutions) to collaboratively train machine learning models without sharing raw data. This approach naturally suits distributed environments. However, several open challenges remain in ensuring that Federated Learning systems meet the broader goals of responsibility and efficiency, including handling non-independent and identically distributed data, energy efficiency, and model explainability. To address these challenges, this work explores four central research questions concerning the evolution of Federated Learning under the lens of Responsible AI. First, it investigates how Federated Learning can be applied across diverse domains and heterogeneous data in a way that supports ethical, privacy-preserving, and human-aligned principles. Second, it examines which techniques and system-level strategies can improve the energy efficiency and overall sustainability of Federated Learning in decentralized environments. Third, it analyzes how explainability methods can be integrated into machine learning workflows to enhance transparency and user trust. Finally, it explores the main challenges that arise when deploying Federated Learning systems in real-world settings, where data, infrastructure, and human factors introduce additional complexity. To answer these questions, the thesis presents a comprehensive methodological and experimental framework grounded in applied research across multiple domains.In the field of telecommunications, the thesis examines the application of Federated Learning for base station traffic forecasting using real-world datasets from cellular networks. A comprehensive evaluation is conducted on multiple model aggregation strategies, client-level personalization methods, and the integration of external data sources to improve prediction accuracy. In addition, the environmental impact of different Federated Learning configurations is assessed through energy consumption metrics, highlighting trade-offs between performance and sustainability. In the healthcare domain, the thesis presents two applications of Federated Learning with integrated privacy-enhancing technologies, targeting anomaly detection and personalized health monitoring. The first application focuses on early-stage identification of Autism Spectrum Disorder using data collected from serious games. A semi-supervised Federated Learning framework is proposed, incorporating autoencoder-based anomaly detection, homomorphic encryption, and advanced aggregation methods to ensure both performance and privacy. The second application introduces the architecture of a decentralized ovulation tracking system that combines Federated Learning, blockchain, and privacy-enhancing technologies at multiple granularities. This system is based on the principle that sensitive reproductive health data remain on user devices while maintaining transparency and trust through immutable consent tracking. To enhance the efficiency and adaptability of decentralized learning systems, a novel client selection algorithm based on Cellular Automata is introduced. This method enables dynamic and context-aware client participation by considering spatial and temporal resource constraints, thus improving the scalability and responsiveness of Federated Learning in resource-constrained environments. Complementing this, the thesis also incorporates a score-based sampling mechanism for smartphones used in a healthcare scenario, where each client is evaluated based on a composite score of local training loss and model divergence. This adaptive selection improves both learning robustness and communication efficiency, particularly under non-iid data conditions. In the context of explainability, the thesis explores explainable machine learning models for detecting non-technical losses in electricity distribution systems. Using consumption data from traditional meters, a combination of model-agnostic techniques, such as SHAP values and permutation importance, is employed to provide transparency and actionable insights to utility providers. Furthermore, the study demonstrates that similar predictive performance can be achieved with a reduced feature set, thus improving both interpretability and computational efficiency. Finally, the thesis addresses the challenges associated with real-world deployment of Federated Learning systems. A critical analysis is conducted on the limitations of current solutions, particularly in terms of statistical heterogeneity, data availability, privacy-preserving mechanisms, and system personalization. Through this analysis, a set of practical recommendations is proposed to bridge the gap between theoretical advancements and operational feasibility. Overall, the thesis shows that building Responsible AI systems is both possible and necessary. By focusing on privacy, explainability, and sustainability, and by evaluating real-world deployment, the work provides practical tools and insights that help make AI more useful, reliable, and aligned with societal requirements.

Degree

thesis:*
Grantor dc:publisher
Democritus University of Thrace (DUTH)
Year dc:date
2025

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Pavlidis, Nikolaos
  • Παυλίδης, Νικόλαος

Subjects

dc:subject × 16

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/60207

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

Pavlidis, Nikolaos; Παυλίδης, Νικόλαος. Optimizing responsible decentralized machine learning methods for AI applications. Democritus University of Thrace (DUTH), 2025. http://hdl.handle.net/10442/hedi/60207