On the robustness of kernel-based pairwise learning