Investigation of the Security Impact of Internet Centralization

Abstract

This thesis investigates how Internet centralization impacts the security of online services, focusing on three problems: digital divides in Australian government services, DNS dependencies of government domains, and privacy in Federated Learning (FL), where each addresses distinct but interconnected aspects of centralization. First, we analyze the DNS dependencies of Australian government domains, identifying potential disparities in service availability between the general population and indigenous communities. By categorizing DNS providers into leading, non-leading, and government-hosted groups, we expose how the digital divide contributes to service unavailability for indigenous domains and increases their vulnerability to outages and attacks. We construct a dataset of Australian government domains to retrieve their DNS providers. Second, we map direct and indirect dependencies through dependency graphs and provide the IP geolocation of DNS providers. We then introduce attacker models by categorizing the attackers' resources and intentions to analyze the implications of DNS dependencies on the vulnerability of different domain groups. Lastly, we address privacy concerns in FL systems, where centralized model aggregation leads to model inversion attacks. We propose ACCESS-FL, a secure aggregation protocol with communication and computation costs as O(1). Experimental results on benchmark datasets (MNIST, FMNIST, and CIFAR) demonstrate that ACCESS-FL significantly reduces computation and communication overhead compared to state-of-the-art methods (Google's SecAgg and SecAgg+) while maintaining comparable model accuracy in honest-but-curious scenarios. This makes ACCESS-FL particularly suitable for large-scale, stable FL environments, such as healthcare systems. In conclusion, this thesis analyses the security consequences of centralization across DNS infrastructure and FL systems to enhance the availability and privacy of online services.