Upsampling Aggregated Network Traffic Data with Denoising Diffusion Probabilistic Models
Nicolas Dupuis, Axel Van Damme, Philippe Dierickx, Olivier Delaby · 2024
In network operations, decisions based on accurate insights are imperative, which heavily rely on data quality. However, availability of high-pacing data is either not practical or asks for extended resources. In this work, we address the above challenge by upsampling the network aggregated traffic data with a Denoising Diffusion Probabilistic Model (DDPM). Our generated realistic-like network traffic traces are consistent with the aggregated volume of bytes measured at the initial pacing and present patterns close from the expected ground truth. The performance of contributed solution is evaluated in an independent Service Quantification (SQ) use case, where the service usage of a subscriber is estimated with improved performances of about 48% compared to the estimation from the initial non-upsampled traffic data.