Variational Autoencoders for Noise Resistant Traffic Generation in B5G Networks

Real-world network traffic data is often a challenge for training deep learning models because of missing values, irregular time intervals, and attacks that can introduce noise or malicious patterns into the data. To overcome this challenge, this paper presents a new technique called Noise Resistant Traffic Generator (NRTG). NRTG incorporates noise reduction methods and constructs a resistant traffic generator capable of processing missing data in realistic scenarios. The proposed NRTG works in two stages. In the first stage, a classifier is used to distinguish between noisy and non-noisy traffic. After filtering the non-noisy traffic, a generator model is used to generate noise-free synthetic traffic samples. For this purpose, a Variational Autoencoder (VAE) is used to generate the synthetic samples and fill in the missing data in the time series. VAEs, as generative models, can capture the underlying data structure and generate plausible missing values. The classification network in NRTG improves the generative model by providing more reliable signals and accelerating convergence. Experimental results highlight the superior robustness and adaptability of NRTG in dynamic and unpredictable network environments.