Paper 2024/1074

Trust Nobody: Privacy-Preserving Proofs for Edited Photos with Your Laptop

Abstract

The Internet has plenty of images that are transformations (e.g., resize, blur) of confidential original images. Several scenarios (e.g., selling images over the Internet, fighting disinformation, detecting deep fakes) would highly benefit from systems allowing to verify that an image is the result of a transformation applied to a confidential authentic image. In this paper, we focus on systems for proving and verifying the correctness of transformations of authentic images guaranteeing: 1) confidentiality (i.e., the original image remains private), 2) efficient proof generation (i.e., the proof certifying the correctness of the transformation can be computed with a common laptop) even for high-resolution images, 3) authenticity (i.e., only the advertised transformations have been applied) and 4) fast detection of fraud proofs. Our contribution consists of the following results: - We present new definitions following in part the ones proposed by Naveh and Tromer [IEEE S&P 2016] and strengthening them to face more realistic adversaries. - We propose techniques leveraging the way typical transformations work to then efficiently instantiate ZK-snarks circumventing the major bottlenecks due to claims about large pre-images of cryptographic hashes. - We present a 1st construction based on an ad-hoc signature scheme and an and-hoc cryptographic hash function, obtaining for the first time all the above 4 properties. - We present a 2nd construction that, unlike in previous results, works with the signature scheme and cryptographic hash function included in the C2PA specifications. Experimental results confirm the viability of our approach: in our 1st construction, an authentic transformation (e.g., a resize or a crop) of a high-resolution image of 30 MP can be generated on a common 8 cores PC in about 41 minutes employing less than 4 GB of RAM. Our 2nd construction is roughly one order of magnitude slower than our 1st construction. Prior results instead either require expensive computing resources or provide unsatisfying confidentiality.