Photo DNA

Abstract

Awareness of sexual abuse of children has grown enormously over the past two decades - however, the recent advancements in technology has also made it easier to propagate it. To tackle this problem, we develop an image fingerprinting technique, which will be invariant to minor alterations in colour, shape, and style. This will help us in tracking down images of child pornography, by matching the fingerprints to a dataset of already identified images. We train a convolutional neural network to learn fixed-length embeddings, such that the geometric, intensity and style transformations of the images have the same embedding. The style transformations are developed using state-of-the-art Generative Adversarial Networks (GANs), while the intensity and geometric transformations use traditional image processing algorithms. This embedding can serve the purpose of a fingerprint, and can be used to uniquely identify any image, even if it is transformed using various techniques.