Adversarially Robust Image/Video Quality Assessment
Benchmarks and defense methods for image/video quality metrics robustness to adversarial attacks.
| Role: Project Head | Period: Apr 2022 – Dec 2024 |
Developed benchmarks of image/video quality metrics robustness to black-box and white-box adversarial attacks. Found vulnerabilities in 15 no-reference metrics and current full-reference SOTA method. Developing adversarial defence methods for image quality metrics, including certification, adversarial training and purification.
Project page: videoprocessing.ai/benchmarks/metrics-robustness.html