Towards Predicting the Success of Transfer-based Attacks by Quantifying Shared Feature Representations