Without model.py, running the attacker script fails immediately. You must act as a
reverse-engineer and read the tensor metadata from the binary file using a simple Python
interactive command string on the terminal:

$ python -c "import torch; b = torch.load('leak.pt'); print({k: v.shape for k, v in b['state'].items()})"

Based on the model parameters of the above output you can recreate model.py:


import torch.nn as nn

class SmallNet(nn.Module):
    def __init__(self):
        super().__init__()

        self.net = nn.Sequential(
            nn.Linear(_, _), 		// insert net.0.weight variables in reverse order [neurons, vecor_length]
            nn.ReLU(),				// net.1 holds no weight matrices or bias vectors, it must be a parameterless non-linear function like ReLU
            nn.Linear(_, _)			// final transformation layer (net.2, variables in reverse order) uses 16 internal dimensions and reduces them down to an output array of length 2 
        )

    def forward(self, x):
        return self.net(x)