Image data is being largely captured by ubiquitous sensing cameras, such as traffic cameras, security surveillance cameras, and body-worn cameras. The publication of image data would greatly benefit the research communities and enable many applications, including the study of driving behaviors toward intelligent transportation systems and the study of building access patterns to improve energy efficiency. However, the publication of high resolution images would raise privacy concerns due to sensitive content captured by the cameras, e.g., identities and activities. Prior to sharing with untrusted parties, the image data must be sanitized to protect the privacy of persons and objects represented in the data.
Recent research studies have shown that images obfuscated via standard obfuscation, such as pixelization and bluring, can be recovered and re-identified (McPherson et al. 2016, Hill et al. 2016). The goal of this project is to develop image data publication solutions with a rigorous privacy notion, i.e., differential privacy, to enable data sharing without compromising privacy.
We propose to adapt the standard notion of differential privacy to image data and aim to develop high utility image obfuscation/sampling algorithms with the rigorous privacy guarantees. The output image, e.g., below on right, should be compatible with existing image processing techniques for accurate extraction of useful information, e.g., blobs detected in green boxes. More technical details can be found in our publications.
Left: Input image; Right: Privacy-enhanced output image
Faculty: Dr. Liyue Fan - Assistant Professor at UNC Charlotte
- Dominick Reilly (UNC Charlotte, graduated May 2020)
- Zachary McVicker (UAlbany, graduated May 2018)
- Dominick Reilly, PhD student in CS, current
- Muhammad Usama Saleem, PhD student in CS, current
- Jeet Jivrajani, MS in CS, UNC Charlotte, graduated May 2020
- Abhilash Mandlekar, MS in CS, UNC Charlotte, graduated May 2020
- Monali Vithalani, MS in CS, UAlbany, graduated May 2019
- Dominick Reilly and Liyue Fan. A Comparative Evaluation of Differentially Private Image Obfuscation. Proceedings of the Third IEEE International Conference on Trust, Privacy and Security in Intelligent Systems, and Applications, 2021. To appear.
- Liyue Fan and Akarsh Pokkunuru. DPNeT: Differentially Private Network Traffic Synthesis with Generative Adversarial Networks. Proceedings of the 35th Annual IFIP WG 11.3 Conference on Data and Applications Security and Privacy (DBSec), 2021. Link Presentation
- Liyue Fan. A Survey of Differentially Private Generative Adversarial Networks. In Proceedings of the 2020 AAAI Workshop on Privacy-Preserving Artificial Intelligence, 2020. Link
- Liyue Fan. Practical Image Obfuscation with Provable Privacy. Proceedings of the IEEE International Conference on Multimedia and Expo (ICME), 2019. Link
- Liyue Fan. Differential Privacy for Image Publication. The 2019 Theory and Practice of Differential Privacy Workshop (TPDP), 2019. Link
- Liyue Fan. Image Pixelization with Differential Privacy. Proceedings of the 32nd Annual IFIP WG 11.3 Conference on Data and Applications Security and Privacy (DBSec), 2018. Link
- UCSD Pedestrian Traffic Database
- MOT 2016 Benchmark
- VGG Face Dataset
- Newly added! Link
- Coming soon!