Visible but imperceivable

Summary

As AI-generated content becomes ever more common in our digital world, the challenge of differentiating between human- and AI- produced images rises both in importance and difficulty. Previously, this has been done using image watermarking and steganography which have been used as a way of embedding within the pixels details of an image’s history or context. Traditionally, steganography techniques have balanced robustness and visibility. Actively changing the visible content of an image has been seen to be undesirable for a variety of reasons. However, the less visible a watermark is, the more prone it tends to be against removal attacks such as compression or transformation, or even just accidental removal. Compression methods tend to target biological, bottom-up weaknesses in the human visual system, thus removing redundancies in the images that human eyes cannot see, and subsequently removing any watermarks or steganography content hidden within these features. At the same time, optical illusions demonstrate that humans commonly have some cognitive and perceptual weaknesses that render them blind to some image effects. Examples include gradual change blindness (“The Changing Room Illusion” by Michael Cohen) and differences in image saliency and selective attention, including that related to top-down processing and ambigrams.

Aim

To create a methodology and variety of techniques that will produce robust watermarks for images/videos by making watermarks that are easily visible/detectable with the naked eye if there is prior knowledge of their existence, but otherwise imperceptible to the casual viewer. 

Objectives

• To explore and critically analyse what facets of human vision and cognition contribute to an image or image transform being perceivable or imperceivable.
• To produce and a visible-yet-imperceivable watermarking methodology for digital images and a subsequent reliable computer vision/machine learning classification/detection method for the presence of such watermarks. The limits of the methodology should be examined as it may be unlikely that a completely imperceivable yet visible watermark exists.
• To evaluate the robustness of the novel watermarking methods against common removal attacks and/or image processing techniques that disrupt watermarks (e.g. rebroadcast attacks, affine transforms, targeted compression)
• To investigate the relationship between image content and context to imperceivable watermarks i.e. how does the watermark have to change to adapt to different types of images and different contexts such as photographs, artwork, medical images etc.

Candidates

Candidates should have some knowledge of digital image creation/processing, either using existing image editing/generation tools (e.g. Adobe suite, StableDiffusion etc.) or programmatically (e.g. Python, OpenCV). Familiarity with computer vision techniques as well as a background in machine learning would be considered an advantage (e.g. TensorFlow, Keras). The project is also likely to involve user studies in order to quantify perception of proposed watermarking methods. 

Supervisors

Discuss this further with a potential supervisor for this research degree:

Research Themes

Find other Research Degrees in the same theme: