Fig. 4. IG attributions revealing incorrect reasoning and guiding the design of adversarial examples.

a The model correctly predicts the major para product of a typical Friedel–Crafts acylation, but low attribution is given to the para directing –F group. b The model is fooled into incorrectly predicting the para product when the –F is replaced by meta-directing groups. The low attributions given to the directing groups indicate that the model has not learnt their importance.