In this project, we aim to devise a new scheme of adveserial learning which pertrube the input to hidden layers, and not solely the input layer. We will consider either training the network with attacks starting with pertrubing the upper layers where it may be easier for the network to generalize and continuing to lower layers. Another approach will add pertubation in each layer (e.g. by making use of the forward-backward scheme multiple times) in order to mimic the effect of a stronger attack, as previous works show that making use of strong attacks during the adverserial training results in more roubust networks.
Such approachs were suggested before but in a somewhat simplified manner.