image: From biological camouflage to multispectral signature management and adversarial attacks on AI vision, each advance in perception has driven new ways to stay unseen.
Credit: ©Science China Press
A research team led by Professor Yongxiang Liu at the National University of Defense Technology recently published a review article titled ‘The Evolution of Anti-Visual Perception: Biomimetics, Spectral Physics, and Artificial Intelligence’ in National Science Review. The review introduces anti-visual perception as a unifying concept, bringing biomimetic camouflage, multispectral signature management, and adversarial attacks on machine vision into one framework.
In the Himalayas, a snow leopard crouches among snow-covered rocks, its coat blending into the mountainside. A hundred meters away, a blue sheep stops grazing and scans the horizon. This silent standoff captures an ancient contest between seeing and staying unseen—one that has shaped life on Earth for hundreds of millions of years. The same contest has played out in human warfare. In World War I, the British Navy used dazzle camouflage to confuse enemy judgement; in World War II, the Allies built a ghost army out of inflatable tanks and fake radio traffic. As sensing moved beyond visible light into infrared and radar, the contest entered the multispectral era. Now that AI interprets visual data, it has reached the algorithmic level, where a patch invisible to the human eye can make a self-driving car misread a stop sign.
According to the review, nature wrote the first chapter of anti-visual perception. Seven hundred million years ago, cnidarians used opsins for light detection, marking the origin of biological vision. In the Cambrian ocean, Anomalocaris dominated with compound eyes of about 16,000 lenses, and color vision appeared more than 300 million years ago. As vision evolved, so did concealment. Nature's strategies have since inspired technology: polar bear hollow hairs inspired low-observable fabrics; leopard spots and tiger stripes led to digital camouflage; leaf-mimicking butterflies, ink-squirting octopuses, and death-feigning frogs evolved into decoys and smoke screens. More recently, scientists have begun recreating biological micro- and nanostructures, producing aerogel fibers inspired by polar bear hair and electro-responsive photonic crystals inspired by chameleons that reversibly switch from blue to red.
Beyond the visible band, modern imaging spans the electromagnetic spectrum from ultraviolet to microwave, so countermeasures must control how matter interacts with waves in each band. Avobenzone and zinc oxide nanoparticles block ultraviolet light; micro- and nanostructures turn mirror-like reflection into diffuse reflection to weaken lidar echoes. In the thermal infrared, three strategies dominate: replicating background thermal textures, modulating emissivity, and suppressing heat conduction, with microfluidics, phase-change materials, and aerogels each playing a role. In the microwave band, composite absorbing materials, smoothly curved stealth airframes, and reconfigurable metasurfaces lead the way. The biggest challenge is multispectral compatibility, because different bands often impose conflicting requirements. Multifunctional stacking and hybrid integration are now bringing full-spectrum invisibility closer to reality.
When the beholder shifts from human to AI, the rules change. Machine vision systems built on sensors and neural networks, from CNNs to DINO and YOLO, now exceed the human eye in certain tasks. But greater capability brings greater vulnerability. Anti-perception has shifted from manipulating physical properties to attacking the decision boundaries of AI. In the digital domain, FGSM applies gradient-based noise, while a one-pixel attack can fool a model by changing a single pixel. In the physical world, adversarial patches can reduce pedestrian detection accuracy by more than 70 percent, and a 3D-printed adversarial turtle is mistaken for a rifle from any angle. Attacks are also spreading across bands: thermal camouflage patches can cut infrared detection accuracy from 95.7 percent to 45.4 percent, and cross-modal patches can deceive visible and infrared sensors simultaneously.
Looking ahead, the authors argue that anti-visual perception is moving from isolated exploration to deep interdisciplinary integration. Biomimetics must move from mimicking form to mimicking function, which requires cross-scale knowledge graphs and scalable manufacturing of multiscale structures. Spectral physics must move from passive trade-offs to active design, combining emerging materials such as graphene and phase-change materials with AI for Science inverse-design platforms. Research on AI needs to bridge the gap between simulation and reality, forging stronger shields with sharper spears.
By tracing the evolutionary logic of anti-visual perception across three threads, biomimetics, spectral physics, and artificial intelligence, the review offers a roadmap for next-generation camouflage and stealth, robust perception, and trustworthy AI.