The concept of "intrinsic blind spots" sounds like from a spy thriller. How can I very slightly alter my face to mess up face recognition software?
The paper describes how to calculate "blind spots" that should hold for various different neural networks (even if they are trained on different data sets).