Empirical Investigation of Deep Learning Approaches for Network Traffic Classification in Nigerian ISP Networks

Accurate network traffic classification enables internet service providers to implement appropriate quality of service policies, detect anomalous traffic patterns indicative of security threats, and optimize capacity allocation for different application categories. Traditional port-based classification has become ineffective as encrypted traffic dominates, making deep learning approaches increasingly necessary. This study empirically investigates the performance of convolutional neural network, recurrent neural network, and attention-based transformer architectures for network traffic classification using packet-level data collected from a Nigerian ISP network in Lagos State. A passive traffic collection infrastructure captured 2.4 million anonymized network flows during a six-week data collection period, with ground truth labels established through a combination of deep packet inspection on unencrypted traffic and controlled traffic generation. Model performance was evaluated across 12 traffic categories including video streaming, VoIP, social media, mobile banking, peer-to-peer, and encrypted web browsing. The transformer-based model achieved the highest classification accuracy of 97.3% across all categories, outperforming CNN at 94.1% and RNN at 92.7%. Mobile banking traffic, due to its distinctive connection patterns, achieved near-perfect classification accuracy across all models. Encrypted traffic categories demonstrated the largest accuracy variance across models and the greatest sensitivity to hyperparameter tuning. The study contributes a publicly shareable Nigerian ISP traffic dataset and evaluates model training requirements against the computational resource constraints typical of Nigerian ISP network operations environments. Keywords: traffic classification, deep learning, ISP network, CNN, Nigeria.