JUNE 2024
|
|
“Side-Channel Analysis against Confidentiality of Embedded Deep Learning Models” by Raphaël Joud Date : 21 june 2024 The usage of Deep Learning (DL) models on embedded systems keeps getting more and more popular. Their security must be ensured as those models may be required to perform sensitive tasks or handle confidential data. This question is notably brought up in European regulation projects. However, securing DL models is not considered at the design phase and their deployment exposes them to physical attacks in addition to numerous algorithmic attacks that already exist. This talk focuses on confidentiality threats of DL models leveraging on physical attacks, especially side-channel analysis. Characteristics behind model performance are targeted with a fidelity objective, meaning that such a scenario aims to obtain a clone rather than just steal model performance. Studies are divided into three distinct topics. First, we study architecture extraction considering quantized models in a restrictive black-box context. Such evaluations have been made using only basic pattern recognition methods applied to circuit electromagnetic emanations. Then, we focus on parameter extraction from DL models embedded in 32-bit microcontrollers (Cortex-M7). Leveraging on an iterative strategy, we highlight several challenges induced by complete model extraction through side-channel analysis. From these results, we consider some countermeasures aiming at strengthening embedded DL model confidentiality.
|
JUNE 2024
|
|
“On jitter transfer in Ring Oscillator” by David Lubicz (DGA-MI) Date : 21 june 2024 Room Amphitheatre Rennes The “jitter transfer principle” states the statistical equivalence between two designs of oscillator based TRNG. A first design consists in two noisy oscillators a first one sampling the other while the second design is a noisy oscillator sampled by a perfectly stable clock signal. This last design is used in particular in [1] in order to compute the entropy rate produced by the thermal noise whereas the first design is commonly used for the implementations and measures. The jitter transfer principle allows to enjoy the good features of the two designs that is being able to measure the jitter and use a stochastic model to compute the entropy rate. A version of jitter transfer principle can be found in [1] but the given formula to translate the statistical parameters of the phase noise does not cover most practical cases and it does not come with the error bounds necessary to certify the computations. In this presentation, we make a more thorough and effective treatment of the jitter transfer principle and give further applications in particular to the computation of the entropy rate of a multiring oscillator-based TRNG. [1] Mathieu Baudet, David Lubicz, Julien Micolod, and André Tassiaux. On the Security280 of Oscillator-Based Random Number Generators. Journal of Cryptology, 24(2):398–425
|
