Statistical fault injection: Quantified error and confidence

Fault injection has become a very classical method to determine the dependability of an integrated system with respect to soft errors. Due to the huge number of possible error configurations in complex circuits, a random selection of a subset of potential errors is usual in practical experiments. The main limitation of such a selection is the confidence in the outcomes that is never quantified in the articles. This paper proposes an approach to quantify both the error on the presented results and the confidence on the presented interval. The computation of the required number of faults to inject in order to achieve a given confidence and error interval is also discussed. Experimental results are shown and fully support the presented approach.