Impact of Deep Learning on Real-Time Network Intrusion Detection: A Quantitative Evaluation Using Benchmark Datasets
DOI:
https://doi.org/10.63125/tnaksd40Keywords:
Deep Learning, Intrusion Detection, Cybersecurity, Network Security, CNN-GRUAbstract
This study quantitatively evaluated the effectiveness of deep-learning architectures for real-time network intrusion detection using heterogeneous benchmark datasets, with particular emphasis on predictive accuracy, attack-specific recognition, computational efficiency, scalability, statistical reliability, and cross-dataset generalization. The analytical framework incorporated NSL-KDD, UNSW-NB15, CICIDS2017, CSE-CIC-IDS2018, BoT-IoT, and ToN-IoT, representing approximately 23.15 million eligible network-traffic observations after preprocessing and an overall retention rate of 94.4%. DNN, CNN, RNN, LSTM, GRU, autoencoder-based, CNN-LSTM, and CNN-GRU architectures were evaluated and compared with conventional machine-learning models. CNN-GRU achieved the strongest binary detection performance, with 99.18% accuracy, 99.06% precision, 99.24% recall, 99.09% specificity, 99.15% F1-score, and ROC-AUC of 0.996, while limiting false-positive and false-negative rates to 0.91% and 0.76%, respectively. The model exceeded XGBoost, the strongest conventional benchmark, by 1.72 percentage points in accuracy. Attack-specific results showed F1-scores of 99.54% for DDoS and 99.31% for DoS, although minority attacks remained more difficult to identify, with worm recall declining to 84.63% and its false-negative rate reaching 15.37%. Cross-dataset evaluation revealed a reduction in mean accuracy from 98.96% within datasets to 91.72% across datasets, demonstrating sensitivity to distribution shifts and heterogeneous traffic characteristics. Computational evaluation showed that CNN-GRU processed approximately 13,240 flows per second with baseline inference latency of 1.46 ms. At 250,000 flows, accuracy remained 98.46% and F1-score remained 98.31%, indicating substantial scalability under increased processing demand. Statistical analysis confirmed significant differences among principal models, F(4, 36) = 42.87, p < .001, partial η² = .83, while CNN-GRU exceeded XGBoost with a large effect size of Cohen’s d = 1.28. Overall, the findings demonstrated that hybrid deep-learning architectures can provide highly accurate, statistically reliable, and computationally efficient real-time intrusion detection, although minority-class recognition and cross-dataset generalization remain important challenges.


