Adversarial Training for Enhanced Image Recognition Security
This project focuses on enhancing the security of Convolutional Neural Networks (CNNs), against adversarial attacks in image recognition tasks. We explore two adversarial attacks in our project which are Fast Gradient Sign Method (FGSM) and L0 Norm. Through experimentation we showed that mixing adversarial images with normal training data can improve system’s accuracy against adversarial attacks . Additionally, we propose leveraging the model itself to generate adversarial images for improved defense using PyTorch Framework. We demonstrate the effectiveness of these techniques in strengthening the model resilience against attacks and increase its robustness.
Arooj Fatima (2020-EE-152A)
Ali Hussain (2020-EE-168A)
Muhammad Aziz Haider (2020-EE-172A)
Subhan Mansoor (2020-EE-175A)
Dr. Ahsan Tahir (Course Instructor)