6Diffusion: IPv6 Target Generation Using a Diffusion Model with Global-Local Attention Mechanisms for Internet-wide IPv6 Scanning

TL;DR

The paper tackles the challenge of probing IPv6 addresses given the vast space by introducing 6Diffusion, a diffusion-model–based target generator that maps IPv6 addresses into a continuous vector space and jointly learns global and local structure via GLF-MSA. It employs a forward diffusion to diffuse seed activity and a Transformer-based denoiser with DDIM sampling to efficiently generate high-quality candidate addresses, followed by a two-stage alias removal postprocessing step. Empirical results show that 6Diffusion achieves substantially higher hit rate ($r_{hit}\approx 0.4673$) and generation rate ($r_{gen}\approx 0.4666$), and a strong non-alias rate ($r_{non-aliased}\approx 0.9310$), while delivering richer diversity in new prefixes across multiple prefixes lengths. These findings indicate that diffusion-based IPv6 TGAs with global-local attention can significantly improve internet-wide scanning efficiency and discovery of active prefixes. The approach provides a scalable, data-driven framework for learning active-address distributions and generating diverse, non-alias candidates for IPv6 space exploration.

Abstract

Due to the vast address space of IPv6, the brute-force scanning methods originally applicable to IPv4 are no longer suitable for proactive scanning of IPv6. The recently proposed target generation algorithms have a low hit rate for existing IPv6 target generation algorithms, primarily because they do not accurately fit the distribution patterns of active IPv6 addresses. This paper introduces a diffusion model-based IPv6 target generation algorithm called 6Diffusion. 6Diffusion first maps addresses to vector space for language modeling, adds noise to active IPv6 addresses in the forward process, diffusing them throughout the entire IPv6 address space, and then performs a reverse process to gradually denoise and recover to active IPv6 addresses. We use the DDIM sampler to increase the speed of generating candidate sets. At the same time, we introduce the GLF-MSA (Global-Local Fusion Multi-Head Self-Attention) mechanism to adapt to the top-down global allocation pattern of IPv6 addresses and the local characteristics of IPv6 address segments, thus better learning the deep-level features of active IPv6 addresses. Experimental results show that compared to existing methods, 6Diffusion can generate higher quality candidate sets and outperforms state-of-the-art target generation algorithms across multiple metrics.

Figures (6)

• Figure 1: The overall architecture of 6Diffusion, encompassing five key steps: IPv6 Preprocessing, Diffusion Model Training (Forward Process, Reverse Process), Generating IPv6, IPv6 Postprocessing.
• Figure 2: IPv6 Address convert to IPv6 Word Address.
• Figure 3: The GLF-MSA module comprises both Global MSA and Local MSA. The Global MSA operates in a top-down manner attention, while the Local MSA is based on multi-scale hierarchical window attention.
• Figure 4: Remove alias address in two steps: coarse- grained and fine-grained.
• Figure 5: Hit rate and Generation rate of target generation algorithms.