For several years Adobe has touted its Sensei framework for incorporating AI into its image editing tools for more realistic noise reduction, cloning, and object removal. Unfortunately, that effort is also one more reason it’s become harder to detect image fakery. So Adobe Research, along with the University of Maryland, are working on a way to use a sophisticated Deep Neural Network (DNN) to detect several types of image hacking.
Splicing, Cloning, and Object Removal
The team’s system isn’t a general-purpose system for finding all types of manipulation. Instead, it has been trained to detect three of the most common: splicing, the compositing of multiple images; cloning, copying a portion of an image and pasting it over another; and object removal.
One of the big challenges for the team was finding enough test images to train their network. They took the interesting approach of using the COCO database of images that include labeled objects, and using an automated tool to perform combinations of these three manipulations on them. That gave them a much larger training set of data than most previous efforts.
Dual-Stream Design Analyzes Image and Noise
Examples of tampering artifacts — Unnatural contrast in the baseball photo and obvious low-noise area in the second image
Using AI to learn how to recognize certain types of image manipulation isn’t new, but recent advances in noise analysis has allowed this project to incorporate a novel dual-stream network. One stream consists of the RGB (image) data, which is passed through a convolutional network that’s trained to recognize certain visible features, such as unusual contrasts or color shifts. The other stream is essentially a noise map of the image, formed by creating a Steganalysis Rich Model (SRM) of it. That map is passed through a network that’s trained to recognize unusual noise patterns — for example, those created if different portions of the image were captured using different cameras with different sensors or default processing.
Several Ways of Securing Images
The problem of detecting fake images is made particularly hard if only the processed image is available. And there are several cases where very powerful tools already exist. First, RAW files are quite difficult to fake. So getting the RAW file is now a common requirement of many major photo contests. Second, on-camera signing of images is a great way to secure their origin. Many high-end cameras already offer that as an option. Signed images, just like any public-key secured data, can be authenticated by any recipient. Similarly, JPEGs captured by most cameras also have distinctive attributes that are different from those in images created with Photoshop. So having the original JPEG, a RAW file, or a signed image are all ways to validate an image, or to use it as a baseline compared with the suspected version.
The Beginning of an AI Arms Race
When the team evaluated their system against other leading research implementations, it did better on almost every metric in all cases. As with many other fields like object and facial recognition, image manipulation and detection looks like one where machine learning approaches will quickly leapfrog other techniques. Of course, the two sides will also be leaping over each other, as tools for image editing produce more natural results in tandem with manipulation detection software becoming more powerful.