Python code, using pytorch only

Q1. Adversarial Perturbations: Using CIFAR 10 dataset and protocol, (i) you are required to implement FGSM on your own and using a basic CNN architecture with three convolution layers, you need to show the impact of the attack. (ii) Use two deep learning architectures of your choice (justify your choice) and any two attacks (other than FGSM) out of which one has to be either Mask based attack (as discussed in the class) or PGD attack. Using the SVHN dataset, you need to show the impact of the attacks and compare and contrast them. (iii) implement a deep learning architecture to detect adversarial attacks. The detection accuracy should be more than 70% of your attack (more the better). [MM: 100] Q2. Deepfake Detection: Create 100 deepfakes/faceswap from the existing tools of your choice using your own face images. Split this data into fine-tune and test sets (50-50). Finetune your model in Ql(iii) and test on the remaining 50 test samples and report the performance. [MM: 100] Q3. Audio Deepfake Detection: A) Record 1000 hindi and 1000 english sentences in your voice sampled from the given text files. Now use any TTS system with an Indian speaker and generate audios using the text that you have recorded. (Recorded audio will be considered as real and generated audio will be considered as fake). Please mention the TTS system that you have used. B) Use an ASSIST architecture and train the architecture on the recorded and generated audios for spoof detection i.e real vs fake classification. C) Use pre-trained (trained on ASVSpoof2019 LA) AASIST and finetune it on your dataset for spoof classification. D) Compare the performance for above two experiments using accuracy and EER metric