Reference no: EM133295368

Finally, you will all compete in an adversarial attack challenge. You are free to use any adversarial attack procedure to implement untargeted attacks on a 2-layer baseline DNN. You are welcome to implement IFGSM or PGD attacks for example. For this problem you will simply submit your adversarially perturbed test images from the Fashion_MNIST dataset. For each image, & should be less than or equal to 125/255 assuming images are normalized to lie between [0,1]. We will evaluate your attacks on our own 2-layer DNN with the same architecture as in the tutorial link above --- however you do NOT have access to the weights/biases of our DNN. The expectation is that your attacks should leverage transferability of adversarial attacks (recall that we have seen that attacks on one DNN model transfer to others.