Researchers at UCLA have built a mirror that deliberately distorts what it reflects. The “lying mirror” is an engineered surface that transforms any image it reflects into a different, innocuous-looking pattern, concealing the original information. It uses no electronics, no computer, and no moving parts. The work appears in Nature Communications.
The device, built by the group of Aydogan Ozcan at UCLA, is an all-optical system based on structured diffractive surfaces, a concept related to the diffractive neural networks the lab has pioneered. The surface is trained through computational optimization to convert an input image into an output pattern that looks ordinary, camouflaging the underlying data.
The result is a passive analog of encryption. An observer sees something ordinary in the reflected pattern; only someone who knows the transformation or has the matching decoder can recover the original image. Because the system is purely optical, it works at the speed of light, consumes no power, and has no software to compromise.
The experimental implementation used a structured micro-mirror array with tiny mirror elements of precisely varied properties. The team demonstrated the effect at wavelengths of 480, 550, and 600 nanometers, covering the blue, green, and red channels of a color image, and built a broadband version operating across a continuous spectral range to make the deception robust under realistic illumination.
The transformations held up against random image noise, unknown rotations, shifts, and scaling of the input features, which matters for real-world applications where the input cannot be controlled.
Potential applications include defense and security, where covert marking and tamper-proof labeling matter, and entertainment, where images that change meaning depending on the viewer could be used in display technology. The authors frame the work as a demonstration of what structured diffractive surfaces can do for visual information processing rather than a fielded product.
The work extends the diffractive neural networks Ozcan’s group has developed: layered surfaces of microscopic structures trained to perform machine-learning tasks such as image classification and object detection purely through light scattering. A lying mirror turns that idea from computation toward concealment: instead of recognizing what it sees, the surface transforms what it reflects into something unrecognizable. The training is computational, but the trained surface is a static physical object that performs its function instantly, with no energy consumption and no latency beyond the flight time of light.
The system has limitations. Each mirror is trained for one specific transformation, so changing the function requires designing a new surface. The demonstrations used image data; video or arbitrary inputs would require additional engineering. The device is fixed once fabricated, with none of the reconfigurability of a digital system, and robustness was demonstrated for the specific transformations tested, not proven universally.
The security implications cut both ways. Defensively, a surface that conceals information in ordinary-looking patterns could protect sensitive visual data, mark products against counterfeiting, or create authentication features that are hard to forge because they are fabricated physically rather than stored digitally. Offensively, the same capability could hide information from surveillance or embed deceptive signals in imaging systems. The dual-use nature of concealment technology is inherent to the field, no different in kind from encryption or steganography in the digital world.
The work shows that the boundary between computation and optics is continuing to dissolve. A surface that can be trained to transform, conceal, and decode visual information while operating passively at the speed of light processes information without a transistor.
Sources: Li, Y., Chen, S., Bai, B. & Ozcan, A. Lying mirror using structured surfaces. Nature Communications (2026). DOI: 10.1038/s41467-026-76488-2.