Digital cameras have been increasingly incorporated into many facets of modern society over the past ten years, and they are now extensively employed in facial recognition, mobile phones, security monitoring, and driverless cars.
These cameras continuously create large amounts of visual data, raising more questions regarding privacy protection.
Several existing approaches solve these problems by using techniques to hide sensitive information from the obtained photos, such as image blurring or encryption. However, since the raw images are already taken before they are subjected to digital processing to obscure or encrypt the sensitive information, such solutions still risk exposing sensitive data. These algorithms’ computation also consumes more power.
Attempts were made to find answers by modifying cameras to reduce the image quality and conceal readily identifiable personal information. However, these methods forfeit the general image quality for all the objects of interest, which is undesirable, and they are still open to hostile assaults to recover the recorded sensitive data.
In a recent study that was published in eLight, it was shown how to develop an entirely new type of imager that AI developed to accomplish privacy-preserving imaging.
In their work, researchers from UCLA under the direction of Professor Aydogan Ozcan described a smart camera that captures specific categories of desired items while instantly removing other types of objects from its images without needing any digital processing.
The new camera’s structure comprises a series of transmissive surfaces, each of which has tens of thousands of light-wavelength-scale diffractive features.
Deep learning is used to adjust the phase of the transmitted optical beams on these transmissive surfaces, allowing the camera to selectively capture specific types or classes of desired objects while erasing the rest.
The resultant layers from the deep learning-based design (training) are manufactured and combined in three dimensions to create the smart camera. When assembled, the input items from the target classes of objects come in front of it, and when they do, high-quality photos are produced by the camera.
In contrast, when the input objects in front of the same camera correspond to other undesirable classes, they are optically wiped, resulting in non-informative and low-intensity patterns comparable to random noise.
This AI-designed camera never captures the direct images of undesirable groups of objects since their distinctive information is all-optically removed at the output by light diffraction.
As a result, privacy is protected to the fullest extent possible since an adversarial assault that gained access to the camera’s recorded images would be unable to retrieve the data. Since the unwanted object photos are not recorded, this function can help lower the image storage and transmission strain on cameras.
The UCLA study team created this special data-specific camera to experimentally demonstrate it. It was designed to precisely and selectively scan only one class of handwritten numbers and was made via 3D printing.
Terahertz radiation was used to test this 3D-fabricated camera by lighting up scribbled numbers. The smart camera could specifically image the input objects only if they were handwritten digits “2,” while instantly erasing all other handwritten digits from the output images, producing low-intensity noise-like features.
This was possible due to the fundamental design principles behind it. The study team also demonstrated that this smart camera is resilient to such fluctuations in illumination by thoroughly testing it under various lighting conditions that were never included in its training.
This AI-based camera concept can be utilized to create encryption cameras, adding an extra degree of security and privacy protection beyond data class-specific imagery. Such a camera conducts a chosen linear transformation optically just for the target objects of interest due to the use of AI-optimized diffractive layers in its construction.
The original image of the target objects can only be retrieved by those who have access to the decryption key, which in this case is the inverse linear transformation. On the other hand, the information of the other unwanted objects is permanently destroyed since the AI-designed camera all-optically deletes them at the output.
As a result, various unwanted objects are detected using noise-like properties even when the decryption key is added to the recorded images.
The smart camera runs at the speed of light and does not require any external power for its processing, except for the illuminating light. As a result, it is quick, data- and energy-efficient, making it ideal for task-specific imaging applications, concerned with privacy, and constrained in terms of power.
Future imaging systems that use orders of magnitude less computational and data transmission power might be inspired by the fundamental ideas of this diffractive camera design.
Together with Professor Mona Jarrahi, the Northrop Grumman Endowed Chair and the head of the Terahertz Electronics Laboratory at UCLA ECE, Professor Aydogan Ozcan, the Volgenau Chair for Engineering Innovation and associate director of the California NanoSystems Institute (CNSI) at UCLA, led this research.
Graduate students Bijie Bai, Yi Luo, Tianyi Gan, Yuhang Li, Yifan Zhao, Deniz Mengu, and postdoctoral researcher Dr Jingtian Hu, all of the UCLA ECE department, are the other authors of this study.
According to the authors, support was also provided by the U.S. Office of Naval Research and the Department of Energy.
Bai, B., et al. (2022) To image, or not to image: class-specific diffractive cameras with all-optical erasure of undesired objects. eLight. doi:10.1186/s43593-022-00021-3.