You shouldn’t trust me: Learning models which conceal unfairness from multiple explanation methods.

Transparency of algorithmic systems is an important area of research, which has been discussed as a way for end-users and regulators to develop appropriate trust in machine learning models. One popular approach, LIME [23], even suggests that model expla- nations can answer the question “Why should I trust you?”. Here we show a straightforward method for modifying a pre-trained model to manipulate the output of many popular feature importance explana- tion methods with little change in accuracy, thus demonstrating the danger of trusting such explanation methods. We show how this ex- planation attack can mask a model’s discriminatory use of a sensitive feature, raising strong concerns about using such explanation meth- ods to check fairness of a model.