Brevitas PyTorch Model Vs its HW Implementation

[oaramoon]

Hi all,

I am trying the craft adversarial examples for the tfc_2w_2a model. What is odd is that the adversarial examples that fool the original Brevitas PyTorch model do not fool the HW implementation of the model. Anyone can give me a clue why that's happening? Shouldn't the HW implementation of the model have a behavior similar to the original Brevitas model for the same input?

Also, does anyone know if there is a way to simulate model inference for the onnx model that is generated in the end-to-end example, prior to deployment?