> Academic opportunities > PHD positions

Differently supervised deep learning methods for side-channel attacks

Technological challenge: Cyber security : hardware and sofware (learn more)

Department: Département Systèmes (LETI)

Laboratory: Centre d'Evaluation de la Sécurité des Technologies de l'Information

Start Date: 01-09-2022

Location: Grenoble

CEA Code: SL-DRT-22-0273

Contact: eleonora.cagli@cea.fr

Secure components exploiting embedded cryptographic mechanisms, for instance smart cards, may be vulnerable to the side-channel attacks. Such attacks are based onto the observation of some physical features measured during the device activity, such as power consumption, electromagnetic irradiation, execution time? the variation of these quantity may provoke an information leakage. A deep analysis of the leakage may lead an attacker to retrieve sensitive information, for instance the secret keys of the embedded cryptographic algorithms, and so to break the device security. In order to analyze the leakages, which are typically collected as high-dimensional signals big datasets, the deep-learning methods are nowadays a privileged tool. Since 2017, the interest of embedded security researchers toward this topic grows very fast, especially because of the efficiency of these methods in the context of profiled attacks. In this context, the attacker has access to a second dataset, over which he has complete knowledge. This second dataset allows him to perform a preliminary supervised training phase. This context is the most advantageous for the attacker. To setup the attacks on the field, for instance in the context of complex secure systems evaluation, this scenario is not available. A perfect control of the handled data is not always possible, and the attack must deal with a partial knowledge (weakly-supervised attack), or a null one (unsupervised or self-supervised attack). In the wide state-of-the-art concerning non-supervised attacks, machine-learning techniques appeared about ten years ago. In particular clustering methods attracted considerable interest. Recently, the scientific community tries to enhance the non-profiled attacks, taking advantage of neural network-based machine learning techniques. Today, the deep-learning research makes clustering algorithms evolve, for example through ?embedding? techniques. These techniques aim at represent data into a space that enhances certain ?useful? relations among data. Moreover, some particular neural network architectures, the siamese networks, are conceived to solve tasks like the verification one, via a weakly-supervised learning. Finally, the reinforcement learning is a concept that has shown very good performances in several problems resolutions, for example in training machines that became champions in games like GO or chess, and this only exploiting a minimal knowledge during training (self-supervised learning). The goal of this research is to gain competence about the state-of-the-art in unsupervised, weakly-supervised and self-supervised machine learning techniques, and adapt them to the side-channel attacks context, in order to enhance performances of attacks when the scenario does not allow a perfect supervised training phase. The research will particularly focus onto attacks against asymmetric cryptographic algorithms implementations (public key cryptography), in order to formalize performing attack strategies, in relation with the several state-of-the-art countermeasures implementations, and deeply analyze their properties.

See all positions Download the offer (.zip)

Email Bookmark and Share