What if a camera pixel could recognize color, store what it saw, and create a fresh secret key before the image ever left the sensor? This paper builds a device designed to do exactly that. In-sensor computing aims to bring data storage and processing capabilities to sensors.

Instead of sending every piece of image data away for separate processing, the sensor itself begins handling the information. That matters because keeping storage and processing capabilities close to the sensor brings those functions directly into the sensing hardware itself.

The design uses two light-sensitive electrode responses: silver responds to blue light, while gold responds to red light, respectively, in operation. Together, these RRAM devices mimic biological dichromatic cone cells, using their different light responses to perform color perception from incoming light.

The same devices also create a secret key from their high resistance, whose values vary randomly. Because the key is produced during every exposure cycle, it can be reconfigured instead of staying fixed. The light response comes from a metal-surface effect that helps break apart tiny conducting paths inside the device, changing how easily electricity moves through it.

Think of each pixel as a two-lock door: two stand-alone devices respond to two light wavelengths, providing the pair of light-sensitive channels. After the image is captured, the devices are reset to high resistance so the next resistance reading can provide the basis for key generation.

Those resistance states contain intrinsic randomness, producing a reconfigurable PUF key that can be used to encrypt the captured image. The two devices respond to light wavelengths of 400 and 600 nanometers, and together they form one pixel. The arrangement is similar to a color image sensor with separate color-sensing parts.

The operating sequence is simple in principle: read the light-driven current to capture the color image, reset the devices, and read their high-resistance states to generate the reconfigurable key. The key does not come from the image itself. It comes from the devices’ random high-resistance values, so generating the key is separate from capturing the image.

The two resistance readings from each pixel are compared with a ten-megohm reference resistance before the binary key is generated. Each pair becomes two binary bits: both readings above the reference produce eleven, both below produce zero zero, and a mixed pair produces zero one or one zero.

The reported tests found a balanced distribution of bits, strong differences between generated keys, and high stability across repeated checks. The device captures a color image and turns each pixel’s light response into bits, while separately harvesting its changing high-resistance state as a reusable secret key.

That key encrypts the image and, when applied again, restores the original. In the demonstration, the devices captured an image by comparing their light-driven current with their dark current. The resulting color information was converted into a two-bit pattern for each pixel.

Afterward, the devices were reset, and their high-resistance readings became the key. The image and key were combined with an exclusive-or operation to encrypt the image, while an authorized receiver used the same key to decrypt it.

The device array therefore captured a color image and generated a key at the same time, balancing image security with efficient operation. The result is a working prototype in which light controls the device’s electrical switching, allowing image capture and color recognition while also producing reconfigurable random keys.

Because the design encrypts both image brightness and color information, it offers a possible route toward protecting more than one kind of visual data. For applications that handle visual data, in-sensor cryptography could connect the sensing terminal with security hardware and strengthen security protocols.

The device combines color sensing with changing, built-in randomness for image encryption. That could keep sensitive pictures protected at the point where they are captured, rather than sending raw information elsewhere first.