Fooling SHAP with Stealthily Biased Sampling
Published in ICLR, 2023
This paper demonstrates the possibility of cherry picking the data samples provided to SHAP in order to change the global feature importance. The specific use-case presented concerns a model audit where a company has to convince an auditor that the disparities in model outcome among protected subgroups are caused by meritocratic features.
Recommended citation: Laberge, G., Aïvodji, U., Hara, S., Marchand, M., & Khomh, F. (2023, May). Fooling SHAP with Stealthily Biased Sampling. In The Eleventh International Conference on Learning Representations.