The problem of role mining, a bottom-up process of discovering roles from the user-permission assignments (UPA), has drawn increasing attention in recent years. The role mining problem (RMP) and several of its variants have been proposed in the literature. While the basic RMP discovers roles that exactly represent the UPA, the inexact variants, such as the δ-approx RMP and MinNoise-RMP, allow for some inexactness in the sense that the discovered roles do not have to exactly cover the entire UPA. However, since data in real life is never completely clean, the role mining process is only effective if it is robust to noise. This paper takes the first step towards addressing this issue. Our goal in this paper is to examine if the effect of noise in the UPA could be ameliorated due to the inexactness in the role mining process, thus having little negative impact on the discovered roles. Specifically, we define a formal model of noise and experimentally evaluate the previously proposed algorithm for δ-approx RMP against its robustness to noise. Essentially, this would allow one to come up with strategies to minimize the effect of noise while discovering roles. Our experiments on real data indicate that the role mining process can preferentially cover a lot of the real assignments and leave potentially noisy assignments for further examination. We explore the ramifications of noisy data and discuss next steps towards coming up with more effective algorithms for handling such data.