Logo image
Back
Understanding new threats against digital forensics and computer vision systems
Dissertation   Open access

Understanding new threats against digital forensics and computer vision systems

Xinwei Zhao
Doctor of Philosophy (Ph.D.), Drexel University
May 2022
DOI:
https://doi.org/10.17918/00001274
pdf
Zhao_Xinwei_202210.80 MBDownloadView
PDF Open Access Open Access (License Unspecified)

Abstract

Anti-Forensics Neural networks (Computer science) Convolutions (Mathematics)--Computer programs Generative aversarial networks Multimedia forensics Physical domain attack Forensic Sciences
Image editing softwares, such as Photoshop and GIMP, allow people to easily modify, remove or create digital contents. Technological advances, however, also make it more challenging than ever to ensure the authenticity and integrity of the multimedia contents. In the past decades, multimedia forensic techniques have been developed to identify sources of digital contents or to detect forged contents. Along with study of multimedia forensics, anti-forensics also attracts a lot of attention from the community. Anti-forensic techniques attempt to find loopholes of forensic algorithms and fool them. Understanding the weakness of the existing forensic algorithms help researchers and investigators to improve forensic techniques. This thesis focuses on understanding the threat, potential, and limitation of malicious attacks against multimedia forensics and computer vision systems. In the first part of the thesis, we investigated the traditional anti-forensic techniques built on linear mathematical models, and discussed the strength and limitation of these attacks. Next, we proposed the state-of-the-art GAN-based anti-forensic attacks against deep learning based forensic algorithms. Additionally, we showed that we can use GAN-based anti-forensic attacks to remove forensic traces from AI-synthesized images to fool CNN detectors. In the last part of the thesis, we extended our research to investigate attacks against computer vision systems in the real world. In particular, our study focuses on defending against the real world attacks that operate by physically modifying an object in a way that does not seem malicious to a human, but can fool computer vision systems. We demonstrated the e effectiveness of our proposed defense against multi-sticker attack attacked US traffic signs

Metrics

47 File views/ downloads
50 Record Views

Details

Logo image