Abstract

The widespread adoption of Internet of Things (IoT) devices has raised the potential of BotNet attacks, which pose major threats to network security. This study dives into the implementation of deep learning and transfer learning models to detect and categorize such attacks on real-world network traffic. We compare the performance of four popular architectures on image-based representations of the Bot-IoT dataset: VGGNet-16, ResNet50, MobileNetV2, and InceptionV3. To evaluate robustness under adversarial settings, adversarial samples were generated using the Fast Gradient Sign Method (FGSM), Basic Iterative Method (BIM), and Projected Gradient Descent (PGD). Furthermore, Generative Adversarial Networks (GANs) were used to test the models with realistic and deceptive inputs. The findings of the experiments show that although all models were highly accurate under normal circumstances (e.g., 99.89% for ResNet50 and 99.93% for VGGNet-16), they differed greatly in how resilient they were to adverse data. With a robust score of 91.78%, VGGNet-16 outperformed InceptionV3 in terms of adaptation to attacks. ResNet50 and MobileNetV2, on the other hand, demonstrated significant performance decreases when subjected to adversary influence. These results underline the significance of adversarial training and robust model selection for IoT security applications. This research offers useful information for creating more robust deep learning-based Botnet detection systems by highlighting the advantages and disadvantages of various topics. The findings also highlight the gaps of ample research on adaptive defenses that may change to meet new threats in
intricate IoT ecosystems. 

Keywords: Adversarial attack, Generative adversarial networks, Transfer learning, Deep learning, Cybersecurity