From what I heard from fastai videos, deep learning ideally should handle imbalanced data, that being said.
Assuming that 100k images are representative of your real world data set, you can try the autoencoder approach used to detect data drift, to identify the anomalies.
1 Like