PhysDiff-Net: physics-guided diffusion for synthetic wireless traffic generation

Synthetic network traffic for benchmarking wireless optimisation algorithms must satisfy statistical, temporal, and physical requirements that existing generative models do not meet simultaneously. To address this challenge, we propose PhysDiff-Net, a conditional denoising diffusion probabilistic model for synthetic WiFi network traffic generation. Employing a 6-layer Transformer encoder backbone (5.62M parameters), the model learns to iteratively denoise network traffic data conditioned on hour-of-day labels. We introduce three components: (1) Variational Hour Conditioning with an auxiliary classification loss that enforces strong temporal pattern encoding, enabling the generation of distinct traffic profiles corresponding to peak and off-peak periods; 2) Physics-Guided Training Losses that incorporate an SNR–user-density anti-correlation as an explicit wireless-domain prior, enabling the generator to represent interference-limited high-load conditions that are under-represented in the observational dataset; and (3) Variance-Guided Diffusion Sampling (VGDS), a sampling-time intervention that resolves the trade-off between the strong classifier-free guidance required for temporal steering and the preservation of realistic output variance. VGDS applies an adaptive post-hoc correction that constrains the output distribution to match real data statistics while preserving temporal patterns and physics correlations. PhysDiff-Net achieves a 100% Pattern Score, satisfying all three physics validation criteria, and preserves cross-feature correlation structure (ρFC = 0.856), outperforming VAE, TimeGAN, and vanilla diffusion baselines on five of six evaluation metrics.