A Study on Dataset Development and Model Vulnerability to Backdoors

Abstract

Intrusion Detection Systems (IDS) are crucial components designed to detect and prevent unauthorized access to network resources. In this research, we implemented an AI-based IDS through a multifaceted approach that included creating a custom IDS dataset, evaluating it using a Convolutional Neural Network (CNN) model, and analyzing the security and resilience of the CNN model against backdoor attacks. The experimental results demonstrated a significant improvement in the model's accuracy and its resilience to certain types of attacks. However, vulnerabilities to backdoor attacks were still present. Specifically, the successful insertion of hidden triggers into the CNN model during the training phase revealed the model's susceptibility to these types of attacks. These findings emphasize the urgent need for improved strategies to mitigate backdoor attacks in the design and implementation of IDSs.