First off, awesome to hear that there are new grad courses on this topic. Sounds like the course focused more on pre-processing techniques, which can be helpful but agree with your point. Just to note, there are also training-time modeling techniques that can help with model bias depending on the application (TensorFlow Model Remediation for example).
(Disclaimer I helped to build TensorFlow Model Remediation)
They used a shaped charge to knock off the detonators from the larger payload. Basically used a smaller, targeted explosion to disable a larger explosion.
(Disclaimer I helped to build TensorFlow Model Remediation)