Abstract
Deep learning has become extremely popular and widely adopted across various fields, such as healthcare, business, and social media, over the last couple of years. Convolutional neural networks (CNNs) have proven their values in a variety of visual recognition tasks, but with the increased complexity of the data involving topological relationships, graph neural networks (GNNs) were introduced. Nonetheless, from intriguing observations, it has been found that CNNs are vulnerable to small adversarial samples, and GNNs can be compromised by minor modifications in graph edges without altering the overall structure. These actions can be taken in less noticeable manners and thus bring up significant security issues in real-world applications. To explore the roles of graphs in adversarial samples and learning for social good and security, this thesis studies attributed networks and unsupervised learning and present novel graph-induced algorithms. This work first studies the data leakage issue on the attributed network and develops an algorithm by leveraging the positive aspects of the adversarial attack methods. This approach involves perturbing nodes together with the graph structure with unnoticeable changes, acting as a countermeasure to data hacking. To this end, this dissertation investigate the robustness of unsupervised learning, especially graph-based clustering and develop a generalized framework to exploit the vulnerable aspects of unsupervised learning and its impact on privacy and data sharing over distributed systems. Additionally, this thesis examines the impact of adversarial examples in the context of multi-label classification and develops a graph-based defense algorithm. This dissertation provides a comprehensive examination of adversarial attacks in both supervised and unsupervised learning, focusing on graph-based approaches. It proposes defense algorithms to address issues such as privacy leakage and data hacking, aiming to enhance the security and reliability of machine learning models in real-world environments.