This repo contains the code for our paper titled "an algorithm for out of distribution attack to neural network encoder"
https://arxiv.org/abs/2009.08016
Python: 3.7
Pytorch: 1.5
Currently, using the OOD Attack, we tested 13 OOD detection methods (Classfication-based and Glow-based) reported at ICLR, NeurIPS, etc, and the AUROC scores of these methods are close to 0.5 (random guess).
The code, data and result are in a zip file shared via Google Drive. You can download it (~20GB) from
https://drive.google.com/file/d/1LVDuWeIca_-aQ4E_YGFgZeoq2e3E3yLB/view?usp=sharing
A subset of the code is in the "master" brach of this repo, which contains a demo in the file: app/BS/demo.ipynb
In another work, we show that reconstruction-autoencoder based OOD detection was ineffective for a medical image analysis application.
The paper is here
https://arxiv.org/abs/2102.02885
The code is here
https://github.com/jiasongchen/SPIE-2021
Conclusion:
AI doctor on a human-expert level still remains to be a distant dream.
Current neural network-based medical systems should serve only as assistants to human doctors.
Send me an email if you have any questions: liang at cs dot miami dot edu
my blog: https://liangbright.wordpress.com