Defending Security and Privacy of Image Data from Learning-Based Threats
- Publication Type:
- Thesis
- Issue Date:
- 2023
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
This thesis delves into the significant security and privacy challenges of adopting learning-based methods in image data. It focuses on two major threats: surveillance and tracking, and malicious forgery and tampering. The research is grounded in developing four key strategies: sensitive information sanitization, forgery media detection, media authenticity protection, and media authorship proof.
The first strategy addresses the challenges of surveillance and tracking by proposing a novel learning-based method for semantically sanitizing face images. This method works in the generative network’s latent space to remove identity information while balancing privacy protection and image utility preservation.
In response to the growing threat of malicious forgery and tampering, the thesis introduces a proactive framework that counters this threat by watermarking face identity features. This innovative detection mechanism surpasses traditional forensic methods, offering a reliable measure of media content authenticity.
The thesis also presents a proactive strategy for media content safeguarding. It involves embedding an invisible watermark into target images, pixel-by-pixel entangled with the image. This watermark acts as an indicator of tampering, allowing the detection of tampered regions by comparing retrieved and original watermarks. This method proves effective against various image tampering techniques, including image copy & move, splicing, and in-painting.
Finally, the research tackles the issue of authorship proof. It introduces a novel method that uses semantic information in images to enhance the robustness of watermarks, ensuring reliable attribution of authorship even under common distortions. This approach not only protects the rights of original content creators but also combats plagiarism and copyright infringement.
Overall, this thesis offers substantial contributions to the field by applying the latest learning-based techniques to address security and privacy concerns in media data. It aims to provide secure and privacy-preserving approaches, contributing significantly to the ongoing evolution of media technologies. The research strikes a delicate balance between fostering innovation and preserving user privacy and security, paving the way for future developments in the domain.
Please use this identifier to cite or link to this item: