Algorithms
This page contains a short description of the
algorithms used in this project, as well as sliding
comparisons of the original and filtered images. If you want to know the full mechanisms, go to the paper or repository using the buttons below each description.
Please pull the slide to see the modified image.
PIRE
Perturbations for Image Retrieval Error (PIRE) is an approach to fooling automatic Content-Based Image Retrieval (CBIR) systems. PIRE prevents the similarity search by filtering the query image at pixel level. That way the feature vector of the query image is changed to not resemble its actual semantics so that the CBIR returns a list of images that are completely different from the query image. Since the filtering happens at pixel level, it will not impact the interpretation of the image content by the human eye.
PerC
PerC is designed to further make the filtered images more distinguishable from their original versions, without sacrificing the protection effects. This is achieved by using perceptual color distance (CIEDE2000) to prioritize the filtering operations on pixels where the color change are less sensitive to the human eye. The resulting images with large yet imperceptible changes also gain higher robustness against practical image processing operations, such as compression.
Logit
In realistic scenarios, it is almost impossible to access the technical details of specific malicious information extraction systems. Logit is specifically designed to address this by also securing substantial protection against systems that are hard to access during designing the filters. The merit of logit is that it does not need intensive resources that are necessary for other related solutions.
