NetDiff: Graph Diffusion with Improved Global Capabilities to Generate and Update Mobile Network Topologies

We introduce NetDiff, a node-conditioned denoising diffusion model that generates directional link topologies and a two-slot transmit/receive parity for mobile ad hoc networks. Directional antennas can yield high throughput but require globally consistent link decisions under sector, interference, connectivity, and half-duplex constraints. NetDiff improves global coherence with Absolute Cross-Attentive Modulation (ACAM) tokens, which provide permutation-invariant global signals and help the model match graph-level counts (e.g., density and sector usage). We also propose partial diffusion to update an existing topology with a small number of denoising steps, enabling fast reconfiguration under mobility. NetDiff reaches over 95 \% of target performance with constant inference time, outperforms heuristic and omnidirectional baselines, and improves over a strong diffusion graph-transformer baseline on key metrics.

Integrating directional antennas into mobile networks is challenging and results in extremely combinatorial problems: to circumvent this issue, we seek to imitate the solutions obtained by a costly solver using a neural network. To do so, we build on top of a discrete node-conditioned graph denoising diffusion architecture to generate and update optimized wireless directional-antenna-network topologies in real time, as well as corresponding communication turns. The link topologies we seek to generate must respect strong structural priors and constraints, while dealing with heterogeneous node layouts. We address this matter by equipping the backbone graph transformer with novel (Absolute) Cross-Attentive Modulation tokens: learned virtual nodes that allow gathering global graph information, with an emphasis on measuring local densities, then modulating the nodes and edges accordingly (while respecting permutation invariance). Such tokens allow the generated graphs to respect the desired structural priors more efficiently. Additionally, in order for future topology changes to be smooth and fast, we propose using partial diffusion to update the links over time. Choosing carefully our noise schedule allows us, under some assumptions, to consider the normalized movement of the nodes as additional noise, homogeneous with the training one. It hence allows us to simply denoise the movement-perturbated topology for a few steps, saving computation time and inducing more continuity between the updated topology and the previous one. Combined with domain-specific features and loss functions, NetDiff provides an expressive and flexible architecture for addressing our challenging combinatorial problem.