Thermal infrared (TIR) sensors are the backbone of safety-critical systems that must operate where visible light fails, such as autonomous night driving, all-weather surveillance, and search-and-rescue. Unlike standard RGB cameras that capture reflected light, TIR sensors detect emitted heat radiation. This physical distinction creates a unique security gap: adversarial attacks designed for the visible spectrum—like printed colorful patches—are often invisible to thermal detectors.
To bridge this gap, researchers have developed physical adversarial perturbations that manipulate heat signatures through passive insulation, active heat generation, and cross-modal material properties.
Manipulating Emissivity and Insulation
The most direct way to fool a thermal detector is to mask or alter the heat signature an object naturally emits. Because TIR sensors map temperature and emissivity to grayscale intensity, creating "cold spots" can disrupt the geometric features a deep learning model uses for identification.
Zhu et al. demonstrated this by creating "Infrared Invisible Clothing." By applying ultra-lightweight aerogel—a highly effective thermal insulator—to specific regions of a garment, the authors created high-contrast "QR code" patterns in the infrared spectrum. These patterns block the wearer's body heat, effectively "blinding" detectors by presenting a thermal signature that no longer resembles a human. Similarly, Wei et al. showed that learnable shapes of insulation material could be strategically placed to hide pedestrians and vehicles from automated surveillance.
Beyond insulation, attackers can exploit emissivity—the effectiveness of a surface in emitting energy as thermal radiation. Zhu et al. developed adversarial car stickers using low-emissivity aluminum. Even if the sticker and the car surface are at the same temperature, the aluminum appears much "colder" to the sensor. The authors found that the heat from a car's engine actually enhances the attack by increasing the thermal contrast between the low-emissivity stickers and the high-emissivity vehicle body.
Active Heat Generation
While passive materials are stealthy and require no power, they rely on existing temperature differentials. Active attacks use hardware to generate heat, allowing for more dynamic or precise perturbations.
Li et al. proved that inexpensive, small light bulbs could be used as heat-emitting components to disrupt pedestrian recognition. By modulating the heat of these devices, attackers can create "hot spots" that mislead TIR detectors. Taking this further, Li et al. introduced AdvDisplay, a reusable array of Thermoelectric Coolers (TECs). This "thermal display" can precisely control local temperatures to create complex adversarial patterns, achieving a 96.1% success rate in real-world environments.
These active methods also enable temporal attacks. Lou et al. developed a physical backdoor attack where a small heat-modulating device acts as a trigger. When the device is activated, the vehicle "disappears" or is misclassified by the model, demonstrating how temperature can serve as a hidden, remotely toggled adversarial signal.
Cross-Modal Attacks and Color-Driven Modulation
Modern autonomous systems often use RGB-T fusion, combining visible (RGB) and thermal (T) sensors for robustness. Attacking these systems requires a "unified" perturbation that is effective in both spectra simultaneously.
One approach is to optimize for shape, a feature shared by both sensor types. Wei et al. and Wei et al. developed unified patches using aerogel. In the visible spectrum, the patch presents a specific geometric shape; in the infrared spectrum, the same material creates a corresponding cold spot by blocking body heat. This geometric consistency ensures the attack is not filtered out by multi-modal fusion algorithms.
A more subtle mechanism involves color-driven thermal modulation. Long et al. exploited the physical relationship between solar absorptivity and infrared radiation. By carefully selecting surface colors, the authors could manipulate how much solar radiation an object absorbs, thereby controlling its temperature and its resulting infrared signature. This allows a standard color-printed patch to function as a dual-modal adversarial tool without requiring specialized thermal-reflective materials.
The Security Gap in Infrared Perception
Despite the critical role of infrared sensors, research into their robustness has lagged behind visible-light studies. Hyun et al. and Niu et al. argue that infrared detectors lack established defense mechanisms, such as robust training techniques tailored to the unique, low-texture characteristics of thermal imagery. As Xiang et al. note, the inability of standard printing to alter infrared signatures previously provided a false sense of security; however, the emergence of material-based and active thermal attacks proves that these systems are highly vulnerable to physical-world perturbations.
Key Foundational Work
Adversarial Attacks for Deep Learning-Based Infrared Object Detection (2021) — contributed 4 concept elements to this direction.
Fooling Thermal Infrared Pedestrian Detectors in Real World Using Small Bulbs (2021) — contributed 3 concept elements to this direction.
Infrared Invisible Clothing: Hiding from Infrared Detectors at Multiple Angles in Real World (2022) — contributed 3 concept elements to this direction.