search

MekAkUActOR / DAFAR-Prototype

DAFAR: Detecting Adversaries by Feedback-Autoencoder Reconstruction --- Prototyping System
https://arxiv.org/pdf/2103.06487.pdf
MIT License
2 stars 0 forks source link

MSTDtcAnomL2 weight? #1

Closed xarryon closed 11 months ago

xarryon commented 1 year ago

thx for this great work! Could u please tell how to get the MSTDtcAnomL2 weights as this weight is very useful to generate threshold. Does the weight is generated by reconsitution for input?

MekAkUActOR commented 11 months ago

Sorry for the late response. Yes you can train the decoder with self-supervised manner (as in the manuscript). L2 in the model name means using l2 reconstruction loss. And I believe the model is also stored in https://github.com/MekAkUActOR/DAFAR-Prototype/blob/main/utils/model/DETECTOR/MSTDtcAnomL2.pth.

But please note that this technique may be already out-dated for adversarial example research. For good defense methods with theoretical guarantee, I recommend you to research on certified training. Please refer to https://www.sri.inf.ethz.ch/teaching/rtai22 for SOTA technologies in this field.