Adversarial attack-based security vulnerability verification using deep learning library for multimedia video surveillance

Abstract

Recently, although deep learning has been employed in various fields, it poses the risk of a possible adversarial attack. In this study, we experimentally verified that classification accuracy in the image classification model of deep learning is lowered by adversarial samples generated by malicious attackers. We used the MNIST dataset, a representative image sample, and the NSL-KDD dataset, a representative network data. We measured the detection accuracy by injecting adversarial samples into the Autoencoder and Convolution Neural Network (CNN) classification models created using the TensorFlow and PyTorch libraries. Adversarial samples were generated by transforming the MNIST and NSL-KDD test datasets using the Jacobian-based Saliency Map Attack (JSMA) method and Fast Gradient Sign Method (FGSM). While measuring the accuracy by injecting the samples into the classification model, we verified that the detection accuracy was reduced by a minimum of 21.82% and a maximum of 39.08%.