Adversarial Machine Learning Using Convolutional Neural Network With Imagenet

Abstract

Adversarial attacks are types of attacks where adversaries try to deceive the machine learning algorithm by providing deceptive input. Adversarial attacks are focused on providing inaccurate data at its training phase, or introducing maliciously designed data to deceive an already trained model. Adversarial instances are an excellent element of security to focus on because they represent a concrete challenge that AI is currently facing, and they are also a challenging component of security to work on because it requires a significant amount of research effort and time. Systems that are currently in use are very vulnerable to adversarial attacks. It is as simple as vandalizing traffic signs for the self-driving car to make mistakes. Any machine learning model is easy to attack, as we can feed them with malicious and wrong data input with ease. This paper focuses on the Adversarial manipulation of the Machine Learning Algorithm and illustrates how attacks are curated towards an Image-based trained model. A series of experiments based on the ImageNet images and pre-trained MobileNet, Convolutional Neural Network(CNN) model from TensorFlow is used to show how an adversarial attack based on Images can be curated to outwit the machine learning model. This paper will primarily focus on the AI/ML algorithm manipulation with one-pixel, multi-pixel, and all-pixel attacks on the TensorFlow pre-trained model.