Integrating cameras and vision algorithms into our daily lives has led to the development of a wide range of new applications but also has raised significant privacy concerns. This thesis reimagines these applications in a privacy-aware fashion, enabling optimal privacy-utility trade-offs. The solutions it explores leverage the design degrees of freedom offered by three domains: optics, electronics, and digital computing.

In the initial segment, the thesis delves into the utilization of optical computing to obstruct facial identity while facilitating downstream applications such as depth estimation, human pose estimation, person detection, and activity recognition. This optical computing is enabled by either a single-layer diffractive optical element or a metasurface whose parameters are optimized in an end-to-end learning pipeline using adversarial optimization. The results are computational cameras that achieve optimal privacy-utility trade-offs and are validated using proof-of-concept hardware.

The subsequent part of the thesis examines the application of an analog electronic chip designed to execute a shallow Convolutional Neural Network (CNN) for per-pixel analysis. This electronic NN is trained to output only pixels with non-private information, avoiding imaging faces. This method underscores the potential of electronic analog computing in enhancing privacy in vision systems.

Finally, the thesis presents a decentralized digital solution that facilitates the collaborative creation of global crowd-sourced Neural Radiance Fields (NeRFs). This involves the introduction of a novel federation scheme and a secure multi-party computation protocol, ensuring high-quality 3D reconstruction for immersive viewing without compromising the users' privacy.