In last ten years, ear recognition has attracted the interest of scientific community. The advantages of this biometric technology include the remote acquisition, permanence in shape and appearance along time and relatively uniqueness for each individual. This paper focuses on the robustness of local descriptors features for ear recognition and includes the evaluation of two promising techniques: SIFT and Dense-SIFT. The experiments include two public available databases as well as synthetic and real occlusion. The obtained results suggest the promising performance of the proposed local descriptors under controlled conditions. Nevertheless, the distortions and the quality of the sample are strongly determined by the level of collaboration of the subjects. In security applications related to surveillance or forensics such collaboration could be null. The results under hard conditions highlight the difficulties of such features in presence of elevate real distortion and the necessity of further improve the traditional approaches.