Abstract
The primary goal of this thesis is to provide scientific material intended to better assess and quantify the threat of physical attacks targeting mobile devices. We first describe the threat model and detail how and why an adversary may want to compromise the security of a mobile device using physical means. If no specific countermeasure is implemented on these platforms against physical attacks, they might become more approachable than most high-severity software exploits. Ultimately, protecting mobile devices against physical attacks is not harder than with smartcards and other security products, as most countermeasures are directly applicable and at least equally efficient. However, the main difficulty remains in accurately evaluating the resistance of such devices, and therefore lies in the development of analysis techniques and methodology. To illustrate this threat and evaluate the feasibility of such attacks, we performed multiple studies targeting various cryptographic primitives integrated in System-on-Chips: an Android library performing RSA; hardware implementations of AES as part of ARMv8 instruction set or as a dedicated cryptoprocessor; and the architecture of a typical implementation of a secure boot on smartphone. Our results demonstrate that the inherent complexity of the device does not contribute to its security, but rather result in an underestimation of the risk. In particular, the side-channel signals measured on System-on-Chips are intricate and very complex to analyze. This does not prevent leakage of sensitive data, but only reduces our ability to extract the available information, and thus our ability to assess the residual risk side-channel attacks represent. To address that issue, we developed a probability Mixture Model, along with a set of Spatial Statistical Analysis techniques, suitable to analyze complex signals of System-on-Chips found in mobile devices. This led to significant improvements over state-of-the-art analysis techniques, which imply that some devices thought to be safe against side-channel analysis might in fact be vulnerable. Indeed, the risk of attack should be re-evaluated with this new capabilities to extract information from jittery and intricate signals, and especially when the implementation relies on Boolean masking and hiding countermeasures such as shuffling.