CLGuard: Presentation Attack Detection Model using Clean Labels for Copyright Protection

Presentation Attack Detection Model (PADM) plays a crucial role in biometric authentication, particularly in fingerprint Liveness Detection Model (FLDM). However, FLDM is susceptible to various unauthorized usage and dissemination threats, highlighting the urgent necessity to protect its Intellectual Property (IP). While previous backdoor watermarking techniques have been effective in authenticating the IP of FLDM, they unfortunately pose a risk to its performance. Thus, this paper presents a novel model watermarking strategy named CLGuard, leveraging robust fingerprint attributes to create authentic label watermarks, facilitating the verification of ownership for FLDM. Notably, these watermarks seamlessly integrate into both the input and output layers, minimizing any irreversible modifications to the feature space of FLDM, ensuring optimal performance. Initially, a selection of High Complexity Inputs (HCI) is chosen based on robust fingerprint attributes, and watermark data is embedded into the HCI set using imperceptible backdoor. Simultaneously, a coding network is deployed to conceal the watermark data within the background layer of the samples, mitigating any adverse impacts on the model's deep feature space. Subsequently, a mapping function is established between triggers and original labels, enabling the fine-tuning of the model with embedded watermarks on a diverse mixed dataset. Ultimately, IP is authenticated by leveraging trigger sets and original misidentification sets. Comprehensive experiments on multiple benchmark datasets demonstrate that CLGuard effectively authenticates suspicious PADM IP without introducing anomalous input-output pairs, thus preserving the original task performance. Furthermore, it withstands attacks such as fine-tuning, neuronal cleanse, and watermark removal, showcasing reliability, stealthiness, fidelity, and robustness.