Non-Invertible Transforms for Image-Based Verification

Imaging may play a unique role in verifying the presence and distribution of warhead components in warhead counting and dismantlement settings where image information content can distinguish among shapes, forms, and material composition of items. However, a major issue with imaging is the high level of intrusiveness, and in particular, the possible need to store sensitive comparison images in the inspection system behind an information barrier (IB). Reducing images via transformations or feature extraction can produce image features (attributes) for verification, but with enough prior information about structure the reduced information itself may be sufficient to deduce sensitive details of the original image. Further reducing resolution of the transformed image information is an option, but too much reduction destroys the quality of the attribute. We study the possibility of a one-way transform that allows storage of non-sensitive reference information and analysis to enable comparison of transformed images within IB constraints. In particular, we consider the degree to which images can be reconstructed from image intensity histograms depending on the number of pixel intensity bins and the degree of frequency data quantization, as well as assumed knowledge of configuration of objects in the images. We also explore the concept of a “perceptual hash” as a class of transforms that may enable verification with provable non-invertibility, leading to an effective one-way transform that preserves the nature of the image feature data without revealing sufficient information to reconstruct the original image.