A Practical Upper Bound for the Worst-Case Attribution Deviations