Bonus Assignment for the ML-lecture for WWI-DSA-18 at DHBW Mannheim

In this Repository I worked on Copycat CNN and Knockoff Nets via their respectiv Trusted-AI implementation in the Adversarial Robustness Toolbox.

In addition because those did not work properly I captured the idea of a copycat cnn in a simple transfer learning approch: I took a trained classifier used by a fellow student and tried if I could perform simular performance by taking a pretrained VGG and using a not problem domain specific dataset (cifar100).

The PDF and the Assigment_9936663.ipynb make up the proposal for the assignment, the rest is further explained below:

Assigment is to be considered the main proposal and contains the simple approch explained in the paper.

Own_Knockoff is a failed approch, which took me much time and effort trying to solve conflicts between PyTorch, Tensorflow 1 and 2 and the used ART Framework. It helped me understand the topic but does as of today not properly work.

Simple is a very simplfied version of Own_Knockoff and was just for me to understand the framework better.

Google Drive Links to Models and Data:

• Bees and Ants
• Victim-Model
• Trained_Copycat VGG
• Trained Copycat Res
• Trained Victim VGG in the Failed Approch