Adversarial Defence
Overview
An adversarial Defence. As an academic project for part of my Advanced Artificial Intelligence module, we had to design an adversarial defence that would be used on the Fashion-MNIST dataset and tested against the other student's attack algorithms. My adversarial defence scored an average accuracy of 59.49% across clean and contaminated data, scoring 67% on the assignment. The reasoning for my defence being weaker than my attack was due to poor training algorithms, not training it sufficiently with contaminated data and over fitting to a particular set of attack algorithms. The iterations of my defence are available on my GitHub.
Tech & Tools
- Python