Modern cameras are not designed with computer vision or machine learning as the target application. Digital imagery collected by cameras can be processed for visual recognition, however the same imagery may also be viewed and interpreted by a human observer. This imagery can reveal much more information than intended, especially for imagery collected in private spaces such as the home. This introduces potential civil liberties violations due to the right to privacy, which requires that subjects have the right to control their private life and keep private issues private. One strategy to address this is to perform visual recognition on the edge, however this approach suffers from (i) cameras that can still leak private infomation if misconfigured or compromised by an attacker, (ii) reduced performance due to power and memory limits for processing on the edge device and (iii) non-existent support for activity recognition.

Encrypted visual sensors are custom optics and vision hardware coupled with a machine learning system called a key-net. This sensor applies a private transformation encoded in the sensor optics/analog preprocessor forming a keyed image. The figure above shows example keyed images analogous to "optical privacy glass" that range from low-privacy to high-privacy. A high-privacy keyed image cannot be interpreted by a human (without knowledge of the key encoded in the sensor optics), but the same image can be interpreted by a paired key-net. The key-net can perform exact encrypted inference on this keyed image with no loss in performance. Our work shows that a keynet sensor is equivalent to optical homomorphic encryption, which is the first practical solution to encrypt visual data at rest, in-transit and in-use. Our goal is to combine key-nets with on-demand consented training sets from Visym Collector to enable the first privacy preserving visual AI system for people in private spaces where no one wants a "camera" watching them.

Our research team is performing the fundamental research to mature this technology (U.S. Patent No. 11,722,641).

---