Unsupervised Federated Learning at the Edge

The rapid advancement of edge computing and the increasing need for privacy-preserving data processing have driven significant developments in distributed machine learning. Federated Learning has emerged as a prominent paradigm in this field, enabling collaborative model training while keeping data decentralized. However, while Federated Learning has shown considerable promise in supervised learning tasks, its application to unsupervised learning remains largely unexplored.This thesis addresses this critical gap by developing novel methodologies for unsupervised federated learning in edge environments. Our research focuses on creating a comprehensive framework for handling unlabeled, distributed data in federated settings, with particular emphasis on addressing the challenges of non-IID (non-independent and identically distributed) data distributions, and operating within the computational constraints of edge devices.To guide and validate our methodological contributions, we explore two fundamental unsupervised tasks: anomaly detection and clustering. These tasks serve as practical use cases, allowing us to demonstrate the efficacy and versatility of our proposed approaches. Anomaly detection, being our initial focus, provided valuable findings that informed the development of our broader unsupervised federated learning methodology. Subsequently, we extended and generalized these concepts to address the more complex task of clustering in federated environments.Through these guiding applications, we iteratively refine our methods to tackle key challenges in unsupervised federated learning, including adaptive model aggregation techniques, and strategies for maintaining model performance across heterogeneous edge devices.First, we begin with a comparative study of centralized versus decentralized approaches to anomaly detection. This preliminary investigation establishes a foundation for understanding the trade-offs between data-centralized and data-distributed learning architectures, examining the strengths and limitations of each approach, and gaining crucial insights that inform the development of our subsequent federated methodologiesBuilding on these insights, we propose an innovative federated anomaly detection method. This approach introduces a preprocessing phase that groups clients with similar inlier (i.e., normal, non anomalous) patterns, followed by federated training of autoencoder-based models within each group. This method effectively addresses the challenge of multi-normal class anomaly detection in federated settings.We extend our methodology to the more complex task of federated clustering. This work generalizes our approach to scenarios where clients have multiple data patterns, demonstrating the flexibility of our framework in handling diverse unsupervised learning tasks.A key feature of our methodologies is the use of locally trained models as compressed representations of client data, enabling efficient client grouping without compromising data privacy. This approach allows us to leverage the strengths of federated learning while adapting to the heterogeneous nature of edge data.We conduct extensive experiments using consistent datasets and settings across our studies, simulating realistic edge computing scenarios. Our results demonstrate that our federated approaches achieve comparable performance to centralized methods while offering significant advantages in terms of privacy preservation and resource efficiency.Furthermore, we provide a comprehensive analysis of the communication costs associated with our methodologies, a critical consideration for practical implementation in bandwidth-constrained edge environments. In this analysis, a trade-off comparison with the centralized counterpart is given.By addressing the challenges of unsupervised learning in federated settings, this thesis contributes to the broader field of distributed machine learning. The findings and methodologies presented in this thesis have potential applications across various domains, including IoT, healthcare, and industrial systems, where decentralized data processing and privacy concerns are paramount.