While adversarial attacks on visible-light (RGB) computer vision are well-documented, a parallel frontier has emerged focusing on the infrared spectrum. Because infrared sensors—including Near-Infrared (NIR) and Thermal Infrared (TIR)—detect heat radiation or invisible light rather than reflected visible light, they require entirely different physical mechanisms to subvert. Researchers are now demonstrating that the "invisible" nature of these sensors can be exploited to create stealthy, real-world attacks that fool autonomous vehicles and surveillance systems while remaining undetectable to the human eye.
The Physics of Thermal Subversion
The fundamental challenge in attacking infrared systems is that traditional printed adversarial patches do not work; a thermal camera sees the temperature of the paper, not the ink. To bridge this gap, researchers have moved from manipulating pixels to manipulating heat signatures and material properties.
One primary mechanism involves active temperature manipulation. Hao Li and colleagues developed AdvDisplay, a "thermal display" made of Thermoelectric Coolers (TECs) that can dynamically change its infrared signature to fool detectors. Similarly, Hui Wei introduced the HOTCOLD Block, which uses inexpensive warming and cooling patches to create high-contrast blocks that disrupt the human heat signature, rendering a person "invisible" to models like YOLOv5.
Passive Stealth through Material Science
More recently, research has shifted toward passive attacks that require no power source. These leverage the concepts of thermal insulation and emissivity.
- Insulation: By using materials like aerogel, which blocks body heat, researchers can create "cold spots" that appear as specific geometric shapes to a thermal sensor. Xingxing Wei demonstrated that optimizing the shape of these insulating patches is more effective than optimizing texture, as geometry is a feature shared across both visible and infrared modalities Unified Adversarial Patch.
- Emissivity: Xiaopei Zhu and Zhanhao Hu showed that low-emissivity materials, such as aluminum film, can be used as Infrared Adversarial Car Stickers. These stickers reflect the ambient temperature of the sky rather than the car's heat, creating artificial thermal contrast that blinds detectors.
Cross-Modal and Stealthy Triggers
The most dangerous infrared attacks are those that are completely invisible to humans but highly salient to machines. Kehuan Zhang developed the Invisible Mask Attack, using infrared LEDs hidden in a hat to project light spots onto a face, fooling recognition systems while appearing normal to bystanders.
Furthermore, the field is moving toward backdoor attacks. Jian Lou demonstrated a Physical Backdoor where a small heating device acts as a trigger; when the device is turned on, the vehicle "disappears" from the AI's perception. This highlights a critical vulnerability: because humans cannot see the trigger, these backdoors are significantly harder to detect than their RGB counterparts.
Defending the Infrared Spectrum
Defending against these attacks is difficult because infrared images typically have lower resolution and less texture than RGB images, making standard adversarial training less effective. Futa Kai Waseda proposed Patch-based Occlusion-aware Detection (POD), which teaches models to explicitly detect and ignore adversarial patches. However, as Jaeguk Hyun notes, there is still a significant "knowledge gap" in securing infrared-based perception for safety-critical applications like autonomous night driving.