Description
With the widespread development of artifical intelligence, Deep Neural Networks (DNN) have become valuable intellectual property (IP). In the past few years, software and hardware-based attacks targetting at the weights of the DNN have been introduced allowing potential attacker to gain access to a near-perfect copy of the victim's model. However, these attacks either fail against more complex architecture or loose in fidelity regarding the targeted model. In this talk, I will present a new attack methodology combining side-channels attacks with software attacks in order to extract more complex DNN. First, I will present the main methodologies from the state of the art, and the comparison that can be made with traditional cryptanalytic attacks. I will then describe how side-channel attacks can help with the current limitations and introduce our framework as well as our experimental results. This talk will be concluded by the presentation of some potentials future work.