Research Projects

The project aims to analyze and review potential adversarial attacks on existing Deep Neural Networks (DNNs), design robust models for their detection, and conduct robustness analysis against these attacks. The project also studies adversarial attacks on deep neural network-based intrusion detection systems and designs robust intrusion detection systems that are resistant to these attacks. Adversarial models can be classified into white box and black box attacks, with white box attacks using internal information about the target system, while black box attacks use no information about the network. The objectives of the project include confidence reduction, misclassification, targeted misclassification, and untargeted mis-classification. Adversaries' capabilities are defined by the information required for achieving these goals, which can be categorized as training data, network architecture, probability confidence, or samples. The methodology includes studying existing adversarial attacks on DNN models, developing techniques to detect them, developing security mechanisms against different attacks, analyzing and proposing techniques to make DNNs robust against pixel attacks, developing new pixel attacks for various image classification models, and designing robust deep neural networks for different application domains. The expected outcome will involve joint publications on adversarial attacks, robust deep network models, and advance pixel attacks, as well as a patent on the design of a robust deep neural network model against pixel attacks. The project also plans to establish applications in image classification, intrusion detection systems, and anomaly detection systems.