IPv6 active address discovery acts as a cornerstone for large-scale Internet measurement and network security assessment. Current strategies face a fundamental dilemma: Exhaustive probing is computationally infeasible due to the colossal \(2^{128}\) address space, while heuristic and reinforcement learning approaches are frequently misled by the widespread presence of aliased prefixes (“black holes”). To transcend these limitations, this paper proposes 6Diffusion, a novel attention-guided generative framework that reformulates address discovery as a denoising diffusion process on a continuous latent manifold. Unlike traditional methods that rely on discrete pattern matching, 6Diffusion employs a Transformer-based denoiser integrated with gated linear function multi-head self-attention (GLF-MSA) to jointly capture global hierarchical allocation semantics and fine-grained interface identifier (IID) regularities. To address the inherent latency of diffusion models, we implement a deterministic denoising diffusion implicit model (DDIM) sampling strategy, accelerating inference by \(40\times\) through step skipping without compromising generative fidelity. Extensive evaluations are conducted under a unified 100K-seed setting, utilizing active targets reliably sourced from the RouteViews and RIPE RIS public measurement projects collected between January and March 2023. Results demonstrate that 6Diffusion establishes a new state-of-the-art, achieving a 42.43% hit rate and a 91.12% non-aliased rate. This significantly outperforms leading baselines such as 6HMap (40.15% hit, 76.91% non-aliased) and AddrMiner (39.68% hit, 70.93% non-aliased), proving that diffusion-based generation offers a robust, alias-resistant, and scalable paradigm for next-generation network measurement.
