Adversarial attack defences for neural network

• Main Author: Singh Kirath
• Other Authors: Anupam Chattopadhyay
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2022
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision
• Online Access: https://hdl.handle.net/10356/157133
• Institution: Nanyang Technological University

Since the advent of deep learning, we have been wielding them to solve intricate problems in the field of natural language processing, image processing, etc. Furthermore, we have been deploying complex deep learning models in real-time systems like autonomous vehicles, security cameras, etc purely based on their precision only to realize that these high precision models can be vulnerable to a variety of adversaries in the environment, that can hamper the overall robustness of our deep learning models. The contemporary defense strategies in the market either cannot alleviate a variety of adversarial attacks primarily in a white box environment or do not have a standardized approach that can be applied to any form of the complex deep-learning models to make them inert from a variety of adversaries. Moreover, there is a need for standardized adversarial defense strategies for mitigating a variety of adversarial attacks to make our models more robust in a white box environment. In this project, we make use of three different state-of-the-art deep-learning architectures trained on 2 benchmarking datasets – CIFAR-10 and CIFAR-100, to analyze the difference in the performance of these models in the absence of an adversary as well as in the presence of an adversary in a white-box environment. We primarily use two white box attack methodologies – Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) to plant adversarial samples using epsilon values ranging from 0.1 to 0.8. Furthermore, we go one step further to devise a defense strategy – Defensive Distillation, that can be applied to a deep-learning architecture to deplete the overall efficacy of FGSM and PGD attacks.