Deep Learning and Interpretability in Digital Image Forensics

点击翻译按钮获取中文摘要

In the new AI era (or equivalently in the "Fake News" era), seeing will no longer be believing and even truth will not be believed. Ensuring integrity and authenticity of digital images is increasingly challenging and vital. With advanced image editing tools and deep learning (DL) models, people can easily manipulate digital images or generate highly visually convincing fake images and videos, including the infamous DeepFakes, and therefore pose critical challenges in digital image forensics (DIF) analysis. DIF analysis identifies the existence, lack or inconsistency of subtle, perceptually invisible forensic traces in a digital image to validate its origin, integrity and authenticity. Deep learning has been widely employed for DIF tasks. Unfortunately, both digital images and deep learning models are vulnerable to manipulations and attacks, intentionally or unintentionally. Adversarial examples, which an attacker has intentionally designed by adding specific human imperceptible perturbations into clean images, can easily fool a DL model to make a mistake. Particularly with the "black box" nature of current DL models, DL opportunities come with "big challenges" associated with DIF. The nature and scope of the DIF field has rendered deep learning interpretability increasingly critical. With this vision, this proposed research program will focus on exploring the intersection of digital image forensics and deep learning to eventually ensure trusting images and achieve more trusting DL solutions. More specifically, the proposed research program will pursue the following main technical objectives: (a) Exploring deep learning interpretability in digital image forensics: We aim to establish the evaluation benchmark and develop novel interpretability models for both conventional image forensics problems and adversarial deep learning problems; (b) Developing interpretation-incorporated deep learning frameworks for the general-purpose DIF analysis (e.g., simultaneous detection of different image manipulations and attacks); and (c) Developing adversary deep learning approaches to provide model-agnostic adversarial attacks and attack-agnostic adversarial defenses. The proposed research addresses fundamental challenges in digital image security and forensics -- the vulnerability of digital images and deep learning models, and has the potential to make influential contributions in achieving more trusting DL solutions for trusting images. Our expected improvements in the ability to combat fake contents and attacks will be of great benefit to society, privacy and national security. DL is reshaping many industries and revolutionizing many research fields, and our research will help achieve the full vision of trusting DL one tiny step further. The research program will provide an opportunity for the students to be trained in related cutting-edge technologies.