StegaStamp using Pytorch
Contributers: Dan Epshtien and Neta Becker. Based on Jisong Xie's work
Goals:
- Develop a tool that encrypts information into a natural image based on existing code from Berkeley university.
- Improve the algorithm performance in order to receive optimal output images along with good decoding performance.
Notations:
- Original image - the image before encoding
- Encoded image - The image after encoding
- Residual image - the image that is received by substracting the original image from the encoded image (Encoded image - Original image). meaning the values that were added to the original image during the encoding stage.
Results
Using different loss functions, we managed to receive those results
Left to right: residual image, encoded image, original image:
A few more examples (Left to right: residual image, encoded image, original image)
More notations:
- secret loss - The loss function of the encoding
- decipher indicator - Graph that depicts the number of images the decoder managed to decipher out of each batch of 4 images
As seen in the graphs below, there is a trade-off between the two - if secret loss value is low than the decipher indicator is low (meaning we are able to decipher less images out of each batch) and vice versa.