Open TTitcombe opened 4 years ago
Do you mean the colluding adversary? This requires a very different, and much stronger, threat model where the adversary controls the data owner's training process. It is an interesting attack, but in practice, an adversary who can control the training algorithm can do a lot worse harm in most cases than just enabling inference attacks.
Yes, I agree that it would be an unlikely attack in practice. Primarily it's interesting because it demonstrates that overfitting isn't strictly required (the attack works on MNIST), so I imagine it would exhibit a different accuracy loss/membership advantage against epsilon relationship
have I understood it correctly that you only implement Adversaries 1 and 2 of Yeom et al. (in
yeom_membership_inference
)? If so, was there a technical reason the colluding adversary (adversary 3) was not included in your analysis?