| Abstract |
The wide presence of cameras using automatic image processing has a positive impact on smart cities and autonomous driving objectives but also poses privacy threats. In this context, there has been an increasing interest in regulating the acquisition and use of public and private data. In this paper, we propose a method to co-design a lens and an image processing pipeline to perform semantic segmentation of urban images, while ensuring privacy by introducing optical aberrations directly on the lens. This particular design relies on a hardware level of protection that prevents an attacker from accessing sensitive information from the original images of a device. Related works rely on specific privacy threats, requiring a large database for training. In contrast, we propose to seek robustness by preventing deblurring during training in a self-supervised way, thus, without requiring additional annotations. Moreover, we validate our approach by simulating attacks with deblurring, and license plate detection and recognition to show that our model can fool these tasks with success while keeping a high score on the utility task. © 2024 IEEE. |
