Anomaly Detection System for Network Traffic Using Backpropagation Neural Network Optimized with Adam Algorithm

Abstract

Network anomaly detection requires classifiers that distinguish benign traffic from malicious traffic while controlling false alarms. This study evaluates a feed-forward Backpropagation Neural Network (BPNN) trained with the Adam optimizer for binary classification using the NSL-KDD benchmark. The implemented network uses eight retained traffic features, Dense layers of 8 and 15 tanh neurons, and a two-node Softmax output. The supplied implementation uses an 80:20 train-test split with random_state = 42; 20% of the training portion is used as validation. Training uses 100 epochs, batch size 32, learning rate 0.001, and sparse categorical cross-entropy. From the reported confusion matrix (TN=1534, FP=17, FN=23, TP=431), the derived accuracy is 97.99%, precision 96.21%, recall 94.93%, F1-score 95.57%, specificity 98.90%, false-positive rate 1.10%, and false-negative rate 5.07%. An ROC-AUC of 0.992 is retained as reported, but the raw ROC coordinates were not supplied. The results demonstrate strong benchmark performance but do not establish optimizer superiority, zero-day detection, contemporary-network generalization, or production readiness. Controlled optimizer ablation, repeated runs, statistical testing, attack-category analysis, and validation on newer datasets are therefore identified as necessary follow-up work.