Ground Glass Secret Keys Could Deliver High-Performance High-Security Facial Recognition Systems
By scattering light into "speckles" for input into a specially-trained neural network, this cryptosystem works at the speed of light.
A cross-discipline team of researchers has put forward a proposal for a highly-secure facial recognition system which uses a layer of ground glass as a "physical secret key of gigabit length" — and a specially-trained neural network to turn an picture of "optical speckles" back to a recognizable face.
"A scattering ground glass is exploited to generate physical secret keys of gigabit length and encrypt face images via seemingly random optical speckles at light speed," the researchers write in the abstract to their paper. "Face images can then be decrypted from the random speckles by a well-trained decryption neural network, such that face recognition can be realized with up to 98% accuracy."
The idea is to find a fast yet secure means of protecting facial recognition systems from attack. Images from the Flicker-Faces-HQ database are beamed through a layer of ground glass using a spatial light modulator (SLM), with the resulting "speckles" captured by a camera.
To the untrained eye, the resulting image is nothing but noise — but to a neural network, specifically trained on the ground glass filter, the process is partially reversible. While the decoded image is of noticeably lower quality compared to the original, it's good enough for a 98 per cent accuracy rate on facial recognition tests.
It's also effectively impossible to clone the key: "It is nearly impossible to generate the same speckles with a different scattering medium (i.e., the physical secret key), in which the scatterers are randomly distributed, and the propagation behavior of photons is very complicated," the team claims. "Compared with existing digital encryption matrix-based approaches (i.e., relays only on digital secret keys), it is nearly impossible to duplicate the scattering medium to crack the cryptosystem, except for a self-defined medium such as a metasurface."
"The proposed system is fast, low-cost, and easy to integrate with other systems," co-author Puxiang Lai told IEEE Spectrum in a brief interview, which brought the project to our attention. However, it's also extremely difficult to scale: Training the neural network for a single scattering medium took 15 hours on a high-end workstation, a process that would need to be repeated for every device in use and which would generate numerous models — one per sender.
It does, at least, offer a range of advantages over rival approaches, including enhanced security, high performance, and the ability to save on storage space and communication bandwidth: The team found the system worked even when only one-quarter of the speckles were captured and decoded.
The team's work is available under open-access terms on Cornell's arXiv.org preprint server.