Machine learning (ML) has greatly improved intrusion detection in enterprise networks. However, ML models remain vulnerable to adversarial attacks, where small input changes cause misclassification. This study evaluates the robustness of a Random Forest (RF), a standard neural network (NN), and a Transformer-based Network Intrusion Detection System (NIDS). It also introduces ADV_NN, an adversarially trained neural network designed to improve resilience. Model performance is tested using the UNSW-NB15 dataset under both clean and adversarial conditions. The attack types include Projected Gradient Descent (PGD), Fast Gradient Sign Method (FGSM), and Black-Box transfer attacks. The proposed ADV_NN achieves 86.04% accuracy on clean data. It maintains over 80% accuracy under PGD and FGSM attacks, and exceeds 85% under Black-Box attacks at . In contrast, the RF, NN, and Transformer-based models suffer significant degradation under adversarial perturbations. These results highlight the need to incorporate adversarial defenses into ML-based NIDS for secure deployment in real-world environments.
The rapid growth of malicious activities on worldwide communication networks, such as the Internet, has highlighted the need for efficient intrusion detection systems. The efficiency of traditional intrusion detection systems is limited, in part, by their inability to relay effectively relevant information due to their lack of interactive/immersive technologies. In this paper, we explore several network visualization techniques geared toward intrusion detection on small- and large-scale networks. We also examine the use of haptics in network intrusion visualization. By incorporating concepts from electromagnetics, fluid dynamics, and gravitational theory, we show that haptic technologies can provide another dimension of information critical to the efficient visualization of network intrusion data. Furthermore, we explore the applicability of these visualization techniques in conjunction with commercial network intrusion detectors. Finally, we present a network intrusion visualization application with haptic integration, NIVA, which allows the analyst to interactively investigate as well as efficiently detect structured attacks across time and space using advanced interactive three-dimensional displays.
In recent years, ensuring safety in the public domain has posed significant challenges, prompting extensive research into developing intelligent surveillance cameras capable of detecting suspicious items before emergencies arise. Also, it is imperative to mitigate the loss of life and property resulting from terrorist attacks within the public sphere, necessitating effective safety measures. We propose Dual Temporal Buffer Differencing (DTBD) and a Single Short Deep Convolutional Neural Network (SSDCNT) object detector to address these challenges. We built two buffers to model the background independently and uniquely segment the static foreground region. The approach can detect the candidate stationary object using contour methodology with the single shot deep convolutional neural network object detector. The candidate object is validated and identified as abandoned and attended. Our approach is unique because it is robust to various illumination changes without affecting the foreground element, and it can detect suspicious objects in any crowd or complex scenario. Its application will mitigate the risk of a potential terrorist attack in the public domain by sensing and alerting the security expert before a situation becomes an emergency. The performance of our proposed approach, the benefits, and the implementation challenges were evaluated through a publicly available dataset ABODA and PEET 2006. The result demonstrates that our proposed approach outperformed other state-of-the-art algorithms in detecting suspicious items. As a result, the approach can facilitate safety in the public domain. It will enable the security personnel to respond proactively to the suspicious item and de-escalate it before the situation becomes an emergency.