ABSTRACT Due to the massive scale and distributed sparsity of the existing IPv6 address space and the complexity of address allocation patterns, a generative active probing model called Hierarchical Mixture‐of‐Experts Scanner (Hi‐MoE‐Scan) is proposed. It combines structured knowledge guidance and sparse expert modeling, which constructs a basic database through the fusion of multi‐source heterogeneous data, parses IPv6 addresses into five‐level prefixes, and embeds contextual features. A gated sparse expert mechanism is adopted to dynamically select experts for participation in computations, combined with a dynamic capacity adjustment strategy to optimize computational load. Additionally, a cross‐layer self‐distillation mechanism is introduced to enhance the consistency of hierarchical gating distributions. Experimental results show that the candidate address hit rate of Hi‐MoE‐Scan reaches 12.5%, an increase of 28% compared with the existing state‐of‐the‐art methods. The valid rate after deduplication reaches 8.4%, covering more independent active network segments. The average generation time is only 245 milliseconds per 10,000 entries, which is significantly better than mainstream baseline models.
Ma et al. (Wed,) studied this question.