Reference no: EM133284317
AI for Smarter Cyber-Security
Question 1. This part of the assignment is related to the Generative Adversarial Networks (GANs) for image generation.
Select and download a working GAN model (Python code) from the Keras website for the generation of human face images based on the Celebrity Dataset.
1. Run the complete GAN model in a CPU environment. Note the time it takes to fully execute one epoch of the model on the CPU. You must present the screenshots as evidence of the model running on your machine indicating the execution time.
2. Run the complete GAN model in a GPU environment. Note the time it takes to fully execute one epoch of the model on the GPU. You must present the screenshots as evidence of the model running on your machine indicating the execution time.
3. Discuss in your own words the processing time difference between the above 2 processing environments. Which processing environment has performed better and why? How much total time will be needed to train the GAN for 100 epochs on CPU and GPU environments?
4. Provide screenshots of the Generator and Discriminator models of your overall GAN model. Briefly discuss the roles of the Generator and Discriminator models and what type of layers enable them to perform these roles.
Question 2. This part of the assignment is related to Adversarial Machine Learning (AML). Download the following article by Pouya et.al. from the web:
"DEFENSE-GAN: PROTECTING CLASSIFIERS AGAINST ADVERSARIAL ATTACKS USING GENERATIVE MODELS Pouya Samangouei, Maya Kabkab, and Rama Chellappa"
1. Briefly discuss in your own words the Defence-GAN system that the authors have proposed in their article. Provide the relevant block diagram of the system.
2. Briefly discuss in your own words why the proposed system is for 'Defence' purposes and not for Attack Purposes.
3. Briefly discuss in your own words the performance achieved by their proposed system on the MNIST dataset under FGSM Black-Box attack. What factors have you considered in your discussion and why?