Use accuracy only and ignore the minority class.

Train only on the test set.

Delete all fraud cases because they are rare.

Use class weights, resampling, SMOTE, and metrics such as precision, recall, F1, or ROC-AUC.