Towards Trustworthy Biometrics: Generalized Detection of Diffusion-Generated Fingerprint Forgeries in Partial Scenarios

Abstract

Fingerprint recognition is widely used for biometric authentication, yet it is increasingly vulnerable to unauthorized access with the rapid advancement of generative artificial intelligence. In particular, Generative Adversarial Networks (GANs) and diffusion models enable highly realistic fingerprint forgeries, posing serious security threats to fingerprint based systems. Developing deep forgery detection methods that generalize to unseen generative models is therefore critical, especially for partial fingerprints captured by side mounted sensors on smartphones and Artificial Intelligence of Things (AIoT) devices, where limited ridge information constrains discriminative features. Existing approaches often rely on extensive multi generator training or suffer from limited robustness against diffusion based attacks and poor generalization in multi generative model scenarios. In this work, we propose Multi Generative Models Fingerprint Forgery Detection, a generalized framework for partial fingerprint forgery detection that supports both single generative model (SGM) and multi generative models (MGM) training against GAN and Diffusion based attacks. The proposed method introduces four training only modules to enhance feature generalization without increasing inference complexity. By integrating local and global representations, suppressing identity dependent information, and expanding forgery feature distributions through feature augmentation and multi classifier learning, the framework improves robustness to diverse generative attacks, particularly unseen diffusion based forgeries. Experimental results show that under SGM training, the proposed method achieves an average accuracy of 88.57%, outperforming UniversalFakeDetect by 3.33% with a $17.82\times $ faster inference speed. Under MGM training, accuracy improves to 91.11% with two generative models and further reaches 92.47% with three models, while maintaining a lightweight inference time of 0.0161 seconds per image, demonstrating strong generalization across diverse generators and strong potential for edge and AIoT deployment. The source code will be publicly available at: https://github.com/Maohsiuhsu/Towards-Trustworthy-Biometrics-Generalized-Detectionof-Diffusion-Generated-Fingerprint-.git