مشخصات پژوهش

This study proposes an adaptive training framework for Convolutional Neural Networks (CNNs) to address the challenge of long-tail attack classes in imbalanced network traffic datasets. Using a quantitative experimental methodology, publicly available datasets are preprocessed through normalization, feature selection, and traffic representation. A baseline CNN model is first developed using conventional training methods, followed by the introduction of an adaptive approach. The proposed framework integrates adaptive loss weighting, class-aware batch sampling, and curriculum learning to enhance the learning of minority classes while maintaining overall model performance. Independent variables include training strategies and imbalance-handling techniques, whereas dependent variables focus on classification accuracy metrics. The methodology consists of structured stages: data preparation, feature extraction, baseline modeling, adaptive training, and evaluation. Results are expected to demonstrate improved detection of rare attack classes, contributing to more robust and balanced intrusion detection systems in cybersecurity environments.