Fault injection requires the adversary to select good attack parameters (i.e., parameters that will lead to a fault). Finding such parameters can be considered as an optimization problem for which we can apply heuristics. Depending on the fault injection type, we need to consider various parameters, resulting in different search space sizes. For instance, voltage glitching has fewer parameters than, e.g., laser fault injection and can be thus considered as a simpler optimization problem.
This talk first gives an overview of state-of-the-art results with artificial intelligence techniques for fault injection. That includes the usage of evolutionary algorithms, memetic algorithms, and deep learning for voltage glitching, electromagnetic fault injection, and laser fault injection. Next, we discuss how to use a memetic algorithm (combination of genetic algorithm and Hooke-Jeeves local search) for the laser fault injection. Finally, we discuss the role of the initial population (as evolutionary algorithms work on a population of solutions) to reach faulty target behavior.
Marina Krček is a PhD candidate in the CyberSecurity Group at TU Delft, The Netherlands. Her PhD project is a collaboration with STMicroelectronics. She is also a member of AISyLab, the Artificial Intelligence and Security Lab led by Dr. Stjepan Picek. She received her Master's degree in Computer Science in 2017 from the University of Zagreb, Croatia. Her research mainly includes applications of artificial intelligence for implementation attacks with a focus on fault injection and side-channel attacks.