DAI-Lab
Improve robustness against JPEG noise
Metrics: During testing, we should convert the image to JPG and add another metric - JPEG-BPP - which reports the resulting bits-per-pixel. This will most likely be a lot lower than the standard BPP.
Training: We should consider creating a layer which (1) selects random 8x8 patches of the image, (2) applies the DCT transform, (3) zeros out the $N$ highest frequency components where larger $N$ indicates more "noise", and (4) applies the inverse DCT transform.
Hi～thank you for releasing it，it is awesome. I have one question, is it impossible for steganography robust to resizing, rotating and other attacks in current? I have done a small test, resizing and rotating output image will make the hidden message can not be extracted. Is there any steganography method can defend some common image attacks?
Hello, thank you for releasing it , and I am running your codes.I have some questions about your paper: 1, How many pictures do you use to train and valid? 2, Why you use the D which is the upper-bound on the relative payload ?I can't understand this parameter. Please tell me the reasons.
@zhangqizky I also got the same result, this method is not robust against resizing, rotating, etc
Hi～thank you for releasing it，it is awesome. I have one question, is it impossible for steganography robust to resizing, rotating and other attacks in current? I have done a small test, resizing and rotating output image will make the hidden message can not be extracted. Is there any steganography method can defend some common image attacks? 试试鲁棒隐写。。。我最近在尝试
