Due to the worldwide lockdown amidst the COVID-19 crisis, which forced companies, individuals as well as network infrastructures to adapt to the new normal of Work From Home (WFH), Internet Service Providers (ISPs) reported about a 60% increase in Internet traffic when compared to with the data before the pandemic. Consequently, network traffic categorization has become a critical topic in computer networking since it is crucial for activities like bandwidth allocation, resource management, intrusion detection, and virus detection. This work attempts to predict network flows’ bandwidth need and duration using the public QUIC dataset. Many applications benefit from knowing the bandwidth demand and duration. They allow for efficient resource allocation, a better quality of service and faster routing. In the past, traditional ML approaches, such as SVM and Random Forest algorithms as well as deep learning algorithms, such as Deep Clustering, have been implemented to solve the problem of network traffic classification. We show that our Attention-Based Convolutional Neural Network outperforms existing approaches which have relied on traditional Machine Learning, Transfer Learning and Multitask Learning techniques. The proposed model achieves accuracy scores of 96% (Bandwidth data) and 94.30% (Duration data) for five-class classification. We observe a significant increase of up to 6% compared to previous results.