ANOMALY DETECTION IN FLOW BASED IoT NETWORK TRAFIC USING CNN,SMOTE AND PCA

Abstract

This research presents an optimized deep learning approach for network flow anomaly detection, combining data preprocessing, feature classification, and neural networks to improve detection accuracy. The methodology employs systematic data cleaning (handling missing values, normalization) and feature selection before applying machine learning classifiers (Random Forest/SVM) for preliminary anomaly identification. These processed features are then fed into deep learning models (CNN/LSTM) to capture complex temporal patterns in flow data. Experiments conducted on network traffic datasets demonstrate that this hybrid approach achieves superior performance compared to conventional methods, with quantitative improvements in both accuracy (99%) and F1-score (99,5%). The study specifically examines how different preprocessing techniques affect model performance and compares various classification-DL architecture combinations. Results indicate that proper data normalization and feature engineering significantly enhances the deep learning model's anomaly detection capability. This work contributes practical insights for implementing machine learning pipelines in network security systems, showing that a carefully designed preprocessing and classification stage can substantially improve deep learning outcomes. The findings are particularly relevant for developing more reliable intrusion detection systems capable of identifying both known attack patterns and novel anomalies.