Training issue

[jli2972]

Hello,

Thank you for sharing the code and data.

I encountered an issue when running train_feature_extractor and test_rogue_device_detection using the provided parameters and datasets, as described in your paper and code.

The training results for train_feature_extractor are attached. Both the training and validation losses are very low—are these values consistent with your experiments?

However, when I used the trained model for test_rogue_device_detection, the resulting ROC curve (attached) showed an AUC of only 0.53, which is significantly lower than expected. According to the paper, the AUC for extractor_1 should exceed 0.9.

Could you please let me know if you modified any training parameters or preprocessing steps during your experiments? Any guidance would be appreciated.

Thanks.