Visible to the public Biblio

Filters: Author is Guilley, Sylvain  [Clear All Filters]
Facon, Adrien, Guilley, Sylvain, Ngo, Xuan-Thuy, Perianin, Thomas.  2019.  Hardware-enabled AI for Embedded Security: A New Paradigm. 2019 3rd International Conference on Recent Advances in Signal Processing, Telecommunications Computing (SigTelCom). :80–84.

As chips become more and more connected, they are more exposed (both to network and to physical attacks). Therefore one shall ensure they enjoy a sufficient protection level. Security within chips is accordingly becoming a hot topic. Incident detection and reporting is one novel function expected from chips. In this talk, we explain why it is worthwhile to resort to Artificial Intelligence (AI) for security event handling. Drivers are the need to aggregate multiple and heterogeneous security sensors, the need to digest this information quickly to produce exploitable information, and so while maintaining a low false positive detection rate. Key features are adequate learning procedures and fast and secure classification accelerated by hardware. A challenge is to embed such security-oriented AI logic, while not compromising chip power budget and silicon area. This talk accounts for the opportunities permitted by the symbiotic encounter between chip security and AI.

de Chérisey, Eloi, Guilley, Sylvain, Rioul, Olivier, Jayasinghe, Darshana.  2016.  Template Attacks with Partial Profiles and Dirichlet Priors: Application to Timing Attacks. Proceedings of the Hardware and Architectural Support for Security and Privacy 2016. :7:1–7:8.

In order to retrieve the secret key in a side-channel attack, the attacker computes distinguisher values using all the available data. A profiling stage is very useful to provide some a priori information about the leakage model. However, profiling is essentially empirical and may not be exhaustive. Therefore, during the attack, the attacker may come up on previously unseen data, which can be troublesome. A lazy workaround is to ignore all such novel observations altogether. In this paper, we show that this is not optimal and can be avoided. Our proposed techniques eventually improve the performance of classical information-theoretic distinguishers in terms of success rate.