I can verify this is true. It is not extremely difficult to build a multinomial regression model and find "adversarial" images for this trained model. Of course

I can verify this is true. It is not extremely difficult to build a multinomial regression model and find "adversarial" images for this trained model. Of course, you're just taking me for my word on this, but you could very well try it yourself. You don't really need any special software outside the tools of SciPy and sklearn for instance if you use python.