This doesn't rescue their claim. If the suggested class imbalance really exists in the training/test sets, the model will preferentially identify whites as criminals.
The claim is that the model is worse at telling black faces apart from each other.
The system is trained to match images of faces, not identify criminals; it's not comparing things to its training set to give a "criminality" score. The training data is just what has taught the system how to extract features to compare. You run an image of an unknown person against your database of known images, and look for a match so you can identify the unknown person.
If the model is just "worse at" black people, it's going to make more mistakes matching to them.
The system is trained to match images of faces, not identify criminals; it's not comparing things to its training set to give a "criminality" score. The training data is just what has taught the system how to extract features to compare. You run an image of an unknown person against your database of known images, and look for a match so you can identify the unknown person.
If the model is just "worse at" black people, it's going to make more mistakes matching to them.