Gabriel Stalder

Master Thesis
Supervisors: Fabian Mentzer, Dr. Shuhang Gu, Dr. Radu Timofte

Towards Image Cryptography with Deep Learning

Encryption of sensible data is today more relevant than ever, especially for images. They are among the most popular multimedia and often contain private data (e.g. medicine). This thesis explores possibilities for image cryptography with deep neural networks using symmetric-key injection. Different methods are proposed to embed keys in an autoencoder architecture. We show how the model can be trained to distinguish between valid and non-valid decryption keys. Afterwards, it is evaluated how good the network can reconstruct images with true keys and how it tries to hide reconstructions with wrong keys. The final models are able to separate correct and false keys. Some of them yield almost perfect reconstruction for valid keys. The non-valid network output hides colours and textures. The thesis is a first step towards image cryptography with deep learning and the proposed models still need optimizing and te