Adversarial attack detection on internet of things networks

Adversarial attacks pose a significant challenge to deep learning (DL) models, which are increasingly deployed in security-critical applications such as Intrusion Detection Systems (IDS) for the Internet of Things (IoT). These attacks also present a significant challenge to deep learning models, which are increasingly used in security-critical applications such as Intrusion Detection Systems (IDS) for the Internet of Things (IoT). This thesis examines adversarial attacks on network traffic and focuses on vulnerabilities, attack methods, and detection mechanisms. First, vulnerability surfaces in the machine learning lifecycle are identified, and examples and potential weaknesses in the literature are evaluated. Second, existing attacks are reviewed, classified, and the ROSIDS23 dataset is introduced. The ROSIDS23 dataset provides comprehensive traffic data for cybersecurity research in ROS-based robotic systems and constitutes a valuable resource for detailed analyses involving multiple attack types. The proposed Reconstruction Error-based Adversarial Detection (READ) method uses four metrics together to detect adversarial attacks by combining existing metrics and a reconstruction error rate metric, achieving high detection rates. Results show that READ enhances IDS performance by significantly reducing adversarial effects and is particularly effective at low perturbation levels. FGSM attacks, due to their simplicity, are more easily detected compared to iterative attacks such as PGD and BIM. The proposed method detects adversarial attacks with a success rate of 92-100%. Experimental results show that integrating READ into IDS increases accuracy by up to 98% and significantly enhances system reliability.