Project Aletheia: Detecting Adversarial Manipulation of Image Data

The objective of the proposed work is to build an image forensics software tool that detects adversarial data manipulation of images. In Phase II a Convolutional Neural Network will be built to identify additions, subtractions, and modifications to objects in images. Novel elements include the architecture element of an attention network, incorporated to aid the Network in focusing in on relevant elements of an image, much like humans focus in on relevant details while ignoring irrelevant or consistent stimuli. The Network will take in images and return a prediction or likelihood that the image has been modified, as well as a prediction of where and what kind of modification has occurred. Types of modifications on which the Network will be trained include splicing, where parts of one image are copied into another; copy-move, where part of an image is copied onto a different part of the same image; and removal, where part of an image is removed and the background is filled in behind it. A ready-for-use software tool encompassing the Network and a user interface will be delivered at the end of Phase II.