Publication

Abstract : In the present emergence of artificial intelligence, deep learning is at the core of the process. For applications ranging from self-driving vehicles to surveillance and security, computer vision has emerged as the workhorse in the area of artificial intelligence. While deep neural networks have demonstrated phenomenal success (often exceeding the capabilities of humans) in solving complex problems, recent studies have revealed that they are vulnerable to adversarial attacks in the form of subtle perturbations to inputs that cause a model to predict incorrect outputs, according to the researchers. Such disturbances are typically too subtle to be seen in photographs, but they fully trick the deep learning models, which are trained to detect them. Adversarial assaults are a severe danger to the success of deep learning in practice, and should be taken seriously. This fact has lately resulted in a significant increase in the amount of money being donated in this manner. The primary goal of this work is to improvise a design feature of attack and its defence mechanism with centralities on the each neighbours and its performance measure. The overall reduction of accuracy for the before attack and after attack with a marginal error of 5% percent is observed for each type of the Neighbours set with matrix size or even list of elements repeatedly.