An article demonstrates how to generate adversaries for convolutional neural networks (CNNs) by manipulating images. The author, Robert, stumbled upon a way to avoid paying a tax on his pet cat by manipulating an image of the cat to appear as a goldfish to the automated system. Robert uses adversarial machine learning techniques to create an unexpected image that fools the pre-trained resnet50 model by adding slight noise to the image’s array values. The article provides relevant code snippets to facilitate the process. Robert succeeds in forcing the model to detect the image as a goldfish and avoids paying the tax.
source update: My Cat Is a Goldfish, so Dont Tax It. – Towards AI