The growing interest in exploring thermal face biometrics is mostly due to the robustness of thermal imagery to face spoofing attacks. However, this robustness lies in the acquisition of thermal properties by the thermal sensor and is limited to presentation attacks.
In this paper, we propose a new type of attack on thermal face recognition systems, performed at the post-sensor level. In the visible spectrum, this attack would be carried out by simply injecting a face image of the claimed identity into the communication channel right after the sensor. However, unlike visible face images that are abundantly available on the web, thermal face images are not easy to obtain.
Therefore, we propose to generate synthetic thermal attacks by converting visible face images into the thermal spectrum. To perform visible-to-thermal spectrum conversion, we use a cascaded refinement network trained using contextual loss. In a scenario where the attacker has prior knowledge about the spoofing countermeasure of the system, we introduce a new loss computed at the local binary pattern (LBP) maps level to fool an LBP-based spoofing attack detection algorithm.
The vulnerability of thermal face biometric systems to the proposed attack is then assessed using two existing baselines of spoofing attack detection. When compared to the challenging presentation attack using silicone masks, the equal error rate has increased from 0.20% to 11.60% and from 2.28% to 58.54% when exposed to the proposed synthetic attack, using the two spoofing attack detection baselines.
