Enhanced Privacy and Efficiency in Machine Learning Through Innovative Paradigms

Abstract

The rapid evolution of communication technology and the widespread use of Internet of Things (IoT) devices have led to an unprecedented increase in data generation. This surge of data, from sources like smartphones, sensors, and networks, drives innovation across multiple sectors, from healthcare to urban planning. However, this rapid growth also introduces significant challenges, particularly in ensuring privacy as personal information is increasingly collected and shared across digital platforms. Balancing privacy protection with effective data utilization has become a key issue in modern machine learning applications.

This thesis addresses these challenges by exploring the potential of Federated Learning (FL) and Zero-Shot Learning (ZSL) as solutions for privacy-preserving data use. Although promising, these techniques still face gaps in safeguarding user privacy while maximizing data utility. The research presented here aims to bridge this gap, developing new methodologies that protect privacy while enabling efficient exploitation of large datasets.

To address the issue of statistical and system heterogeneity in FL, the thesis introduces an Asynchronous Personalized Federated Learning framework (AP-FL), which incorporates model interpolation and a data-free knowledge transfer method to enhance robustness and efficiency. In the context of Video Summarization, it proposes a frame-based aggregation method and a Community-Aware Clustering Federated Framework (CFed-VS), designed to address privacy concerns and manage the complexity of video data.

Further, the research explores Privacy-Enhanced Zero-Shot Learning (PE-ZSL) and Sentinel-Guided Zero-Shot Learning (SG-ZSL), which offer novel approaches for zero-shot classification without direct access to real data. These frameworks protect sensitive data while ensuring effective knowledge transfer, marking significant advancements in secure AI learning environments.

Through these contributions, this thesis advances the state of machine learning by addressing key issues related to data privacy, heterogeneity, and efficiency. The findings presented here not only improve the robustness of FL and ZSL frameworks but also pave the way for future research into privacy-preserving AI technologies.