6Dime: A resolver-centric reinforcement learning framework for IPv6 DNS infrastructure discovery

Client-side DNS infrastructure is a critical part of the Internet, enabling users to access online services and information. To improve its reliability and security, many measurement studies have examined DNS behavior. However, most existing work focuses on IPv4, and IPv6 DNS resolver infrastructure remains largely unexplored. The main obstacle lies in the inherent difficulty of efficiently discovering these resolvers in the vast IPv6 address space. Although various IPv6 target generation algorithms (TGAs) have been proposed, these TGAs are typically optimized to find more responsive IPv6 addresses rather than to reveal resolver infrastructure from a small set of UDP/53 seed addresses. Thus, their effectiveness in discovering IPv6 DNS resolver infrastructure is limited. To address this gap, we present 6Dime, a reinforcement-learning-based framework that combines a hierarchical multi-armed bandit model with controlled DNS measurements to systematically discover IPv6 DNS resolver infrastructure. 6Dime adopts a resolver-centric objective, guiding an IPv6 target generator with rewards derived from recursive resolvers observed at a controlled authoritative domain. To further improve exploration efficiency, 6Dime makes probing decisions hierarchically at the routable prefix and subnet-ID levels and employs expansion strategies to reduce dependence on seed addresses. Experiments on real-world IPv6 networks show that, under a 200 M probing budget, 6Dime discovers 3.4–4×more IPv6 recursive resolver addresses and achieves 2.5–3×higher infrastructure coverage than state-of-the-art IPv6 TGAs.