The Impact of Input Image Size on The Performance of Deep Learning Models Applied in Cybersecurity
Nguyễn Trung Hiếu, Nguyễn Hồng Sơn · 2024
While deep learning networks such as CNNs and RNNs have been successfully applied in fields like computer vision and natural language processing, their application in cybersecurity presents a unique challenge. The data in the field of cybersecurity is quite different from data in these two areas. A solution that transforms network data into familiar image formats for CNNs, similar to those used in computer vision, has emerged in recent studies. These studies have shown that transformation methods can create images of varying sizes, making image size a customizable parameter. Most studies focus on neural network architectures and selecting hyperparameters to achieve the best model performance. However, whether the variation in input image size affects the model's performance and how it does so remains unclear. This paper provides evidence showing that model performance depends on the transformed image size. Experiments were conducted through two case studies applying CNNs to classify cyber attacks using the botnet dataset CTU-13 and NSL-KDD, respectively. The nature of this phenomenon is predicted through analysis based on experimental results. From this, we also propose a method to transform network data into images of the most appropriate size, which has potential implications for improving the effectiveness of cybersecurity measures. The results contribute to perfecting the method of applying deep learning networks in the field of cybersecurity.