Document Type
Dissertation
Date of Award
8-31-2022
Degree Name
Doctor of Philosophy in Computer Engineering - (Ph.D.)
Department
Electrical and Computer Engineering
First Advisor
MengChu Zhou
Second Advisor
Yun Q. Shi
Third Advisor
John D. Carpinelli
Fourth Advisor
Edwin Hou
Fifth Advisor
Xuan Liu
Sixth Advisor
Frank Y. Shih
Abstract
Digital images are a substantial portion of the information conveyed by social media, the Internet, and television in our daily life. In recent years, digital images have become not only one of the public information carriers, but also a crucial piece of evidence. The widespread availability of low-cost, user-friendly, and potent image editing software and mobile phone applications facilitates altering images without professional expertise. Consequently, safeguarding the originality and integrity of digital images has become a difficulty. Forgers commonly use digital image manipulation to transmit misleading information. Digital image forensics investigates the irregular patterns that might result from image alteration. It is crucial to information security.
Over the past several years, machine learning techniques have been effectively used to identify image forgeries. Convolutional Neural Networks(CNN) are a frequent machine learning approach. A standard CNN model could distinguish between original and manipulated images. In this dissertation, two CNN models are introduced to recognize seam carving and Gaussian filtering.
Training a conventional CNN model for a new similar image forgery detection task, one must start from scratch. Additionally, many types of tampered image data are challenging to acquire or simulate.
Meta-learning is an alternative learning paradigm in which a machine learning model gets experience across numerous related tasks and uses this expertise to improve its future learning performance. Few-shot learning is a method for acquiring knowledge from few data. It can classify images with as few as one or two examples per class. Inspired by meta-learning and few-shot learning, this dissertation proposed a prototypical networks model capable of resolving a collection of related image forgery detection problems. Unlike traditional CNN models, the proposed prototypical networks model does not need to be trained from scratch for a new task. Additionally, it drastically decreases the quantity of training images.
Recommended Citation
Shi, Yuxi, "Digital image forensics via meta-learning and few-shot learning" (2022). Dissertations. 1628.
https://digitalcommons.njit.edu/dissertations/1628
