Seek-and-Hide: Adversarial Steganography via Deep Reinforcement Learning.

Area of Science:

• Computer Science
• Information Security

Background:

• Traditional image steganography can only conceal one secret image within a cover image.
• Existing methods lack adaptability in terms of secret placement and scale.

Purpose of the Study:

• To develop an adaptive local image steganography system (AdaSteg) for improved security and capacity.
• To enable multi-secret steganography within a single cover image.

Main Methods:

• Utilizing deep reinforcement learning to adaptively select optimal local patches for secret concealment.
• Converting secret images into encrypted noise patches, similar to adversarial examples.
• Encoding encrypted secrets into local regions of the cover image.

Main Results:

• The AdaSteg system demonstrates superior performance compared to state-of-the-art methods in both security and capacity.
• Successfully enables multi-secret steganography within a single cover image.
• Introduces a novel criterion for assessing local steganography and a specialized dataset.

Conclusions:

• The proposed AdaSteg system offers a more secure and efficient approach to image steganography.
• The developed dataset provides a standardized benchmark for future research in image steganography.