An AI-Powered Smart Camera Achieves Privacy-Preserving Imaging by Only Recording Objects of Interest While Being Blind to Others

Digital cameras have been widely incorporated into our society over the past ten years. They are now extensively employed in facial recognition, mobile phones, security surveillance, and driverless cars. However, the massive amounts of visual data produced through these photos raise growing concerns regarding privacy protection. Some existing approaches address these concerns by using techniques to hide sensitive information from the obtained photos, such as image blurring or encryption. However, since the raw photographs are already taken before digital processing to obscure or encrypt sensitive information, such solutions still risk exposing sensitive data. Additionally, these methods use more energy when they are being computed. Other attempts to solve this issue include utilizing specialized cameras to reduce the image quality to mask identifiable personal information. However, these methods forfeit the overall image quality for all the interesting objects, which is undesirable, and they are still open to vicious assaults to recover the recorded sensitive data.

A UCLA research team facilitated the development of a fundamentally new kind of imager created by AI. Their work on a novel paradigm for privacy-preserving imaging was also emphasized in a recent paper published in eLight. This intelligent camera design captures only the needed items while instantly removing unwanted object types from the photos without needing digital processing. The design comprises a series of transmissive surfaces, each composed of tens of thousands of wavelength-scale diffractive patterns. Deep learning is used to modify the phase of the transmitted optical beams on these transmissive surfaces, allowing the camera only to capture specific classes of desired objects while erasing the rest. Following training, the layers are produced and combined in three dimensions to create the smart camera. When the input items from the target classes of objects come in front of it after it has been assembled, they create high-quality photos at the camera’s output. Other objects that belong to undesirable classes are optically erased, resulting in the formation of non-informative and low-intensity patterns that resemble random noise.

This AI-designed camera never captures direct images of undesirable classes of objects since their characteristics are eliminated at the camera output due to light diffraction. As a result, privacy is protected to the fullest extent possible, and even if a malicious assault gains access to the camera’s recorded photos, they would never be able to retrieve the data. Since unwanted photographs are not recorded, this feature lowers cameras’ data storage and transmission strain. For experimental demonstration purposes, the UCLA study team designed this special data-specific camera to specifically and selectively image only one class of handwritten numbers. This was made via 3D printing. Terahertz waves were used to test this 3D printed camera by lighting up scribbled numbers. The smart camera was effective in selectively imaging the input items for a certain number while instantly deleting all the other writing digits from the output, producing low-intensity noise-like features. This was accomplished by adhering to the fundamental design principles of the device. By choosing to picture one type of fashion goods and omitting other categories from the output, the researchers also exhibited another variation of the same class-specific camera system. The camera’s design was also tested in challenging lighting scenarios that were not part of its training. The results demonstrated how resistant this smart camera is to changes in illumination.

This AI-based camera may also be used to create encryption cameras, adding an extra layer of security and privacy protection in addition to data class-specific imagery. Such a camera optimizes a chosen linear transformation for the target objects of interest thanks to using AI-optimized diffractive layers in its construction. The original image of the target objects can only be recovered by individuals with access to the decryption key. On the other hand, the knowledge of the other unwanted things is permanently destroyed because the AI-designed camera all-optically erases them at the output. As a result, other undesirable objects will only produce noise-like, unidentifiable features when the decryption key is applied to the captured photos. This smart camera runs at the speed of light and does not require any external power for its processing except 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 lesser orders of magnitude of computing and data transmission power can be inspired by the fundamental ideas of this diffractive camera design.

This Article is written as a research summary article by Marktechpost Staff based on the research paper 'To image, or not to image: class-specifc difractive cameras with all-optical erasure of undesired objects'. All Credit For This Research Goes To Researchers on This Project. Check out the paper and reference article 1 and article 2.

Please Don't Forget To Join Our ML Subreddit