Chapter 7. Model Auditing for Fairness and Discrimination
An audit is a process, system of tools, or expected work product to describe the results of a system complex enough that a certain set of outcomes or qualities cannot be guaranteed by those working in the system. This explains why audits are necessary even when good intentions are assumed by all. The scale or complexity of a system makes it difficult to anticipate potential problems or ensure there are none, so an audit becomes part of assisting regulation of a system to keep it performing according to desired metrics and values.
This chapter takes a different approach to the problem of ensuring fairness. Rather than imagining we are the data scientists or modelers doing data analysis or building out a machine learning system, let’s imagine that we have been handed a system and asked to evaluate it. This can happen under two paradigms: white-box auditing and black-box auditing.
If you are not familiar with this terminology, white-box auditing implies that we can see the code powering a model and get into the internals of a system, whereas black-box auditing means that the internals of the model itself are not available to us but we can still run the model.
White-box auditing might sound easy, but remember that just because a model is available does not mean it is easy to understand. For example, white-box auditing is separate from the issue of interpretability. Remember that even researchers who develop deep learning models ...