TRUE: Towards Robust, Unique, and Efficient Model Fingerprint Generation and Verification.
Abstract
While deep neural networks (DNNs) are being widely adopted, they are also increasingly exposed to model stealing risks, posing significant threats to intellectual property. Model fingerprinting, which utilizes intrinsic characteristics derived from the training process to authenticate models, has been proposed as a potential solution. An ideal fingerprint should satisfy robustness, uniqueness, and efficiency. However, existing methods often struggle to balance these three aspects simultaneously. To address this challenge, we propose TRUE, a twostage fingerprint generation and verification framework. In the fingerprint generation stage, we adopt adversarial examples as the fingerprint type and enhance their robustness through a minmax robust optimization strategy. In the fingerprint verification stage, we assign weights to the adversarial examples using a Boltzmann distribution to reflect their uniqueness, and further employ sampling based on these weights to reduce the number of queries and improve efficiency. Extensive experiments conducted on real-world datasets and across different model architectures demonstrate the superiority of the TRUE framework.