Abstract: At a fundamental level, all Internet applications rely on a dependable packet delivery service provided by the Internet routing system. However, measurements have shown that various faults (i.e., physical failures or misbehavior) occur from time to time in the Internet. For physical failures of routers or links, it takes existing Internet protocols 3 minutes on average to converge to alternative paths, resulting in interrupted packet delivery. For misbehavior (i.e., mis-configurations or malicious attacks) that could result in data traffic hijacking, current Internet routing protocols provide little defense. This dissertation studies how to build resilient Internet routing protocols to provide reliable packet delivery despite the faults in the context of Border Gateway Protocol (BGP). First, we evaluate the performance of the existing routing protocols in the face of physical failures, as measured by their ability to continue packet delivery service during routing convergence. Our results show that quickly propagating new reachability information has the most impact on the packet delivery performance. We also show that BGP's update rate-limiting timer is the major contributing factor to the duration of the transient loops. Second, to speed up BGP convergence and improve packet delivery, we propose the Root Cause Notification approach. This approach explicitly signals the failure information, which enables a node to invalidate all the paths that have become obsolete because of the same failure. It reduces the upper bound of BGP routing convergence delay from O(n) to O(d), where n is the number of nodes in a BGP network and d is the network diameter. We also develop a framework that enables us to fill the gaps in analytical results for existing convergence improvement algorithms. Third, we propose a novel semantic checking approach and develop mechanisms for detecting invalid paths in BGP and further identifying the attacker. In this approach, a node accumulates a network topology using the root cause in formation, path information in BGP message, and on-demand queries. A high detection ratio is achieved by comparing the received paths with the paths predicted based on the accumulated topology.