Adversarially Robust 1D-CNN for Malicious Traffic Detection in Network Security Applications

Journal of Cybersecurity and Information Management JCIM 2690-6775 2769-7851 10.54216/JCIM https://www.americaspg.com/journals/show/3728 2019 2019 Adversarially Robust 1D-CNN for Malicious Traffic Detection in Network Security Applications Computer Department, College of Education for Pure Sciences, Wasit University, 52001 Al-Kut, Wasit, Iraq Oday Oday Computer Department, College of Education for Pure Sciences, Wasit University, 52001 Al-Kut, Wasit, Iraq Esraa Saleh Alomari Computer Department, College of Education for Pure Sciences, Wasit University, 52001 Al-Kut, Wasit, Iraq Jaafar Sadiq Alrubaye Ministry of Education, Wasit Education Directorate, Kut 52001, Iraq Oday Ali Hassen While threats in cyberspace are in a state of constant evolution, the use of AI in cyber defense has numerous opportunities and dangers. This paper evaluates adversarial robustness for deep learning networks in network security applications by introducing a novel one-dimensional CNN model for malicious traffic detection. We conducted rigorous end-to-end processing and analysis of network traffic data, using a balanced dataset of 200,000 connections (46.52% benign, 53.48% malicious). Our model architecture includes three convolutional blocks (32, 64, and 128 filters, respectively) with batch normalization and dropout mechanisms (0.3 and 0.2, respectively). We use standardized feature scaling, label encoding for categorical features, and stratified sampling to maintain class distribution integrity. Our proposed approach achieved remarkable performance metrics compared to standard approaches with a 95% AUC-ROC result (15% better than baseline CNN models) and detection rate of 99.99% malicious traffic (compared to 98.5% with standard architectures). The model demonstrates better robustness with only 10 false negatives out of 107,895 malicious samples, a 67% enhancement compared to current state-of-the-art systems. Training dynamics show great stability with minimal overfitting (validationtraining loss difference of only 0.01), indicating good generalization ability. 2025 2025 162 175 10.54216/JCIM.160113 https://www.americaspg.com/articleinfo/2/show/3728