Copyright Protection for Deep Regression Models: A Three-Stage Verification Method

Read full text Summarise with VIRAT
Authors: Yidong Chai, Xinmei Wang, Yi Liu, Xiao Gang Han, Yang Qian, Kun Yuan, Liu Yezheng
Cited by: 0
References: 53

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

The protection of AI models’ intellectual property has become a critical concern in AI governance. The copyright of deep learning models can be protected by applying watermarking and fingerprinting techniques. While valuable, these techniques still have three main limitations. First, although black-box watermarking techniques are more practical than white-box ones, they remain intrusive and can interfere with the model’s training process, potentially leading to performance degradation. Second, black-box watermarking and fingerprinting techniques have primarily been proposed for classification models. The significant differences between regression and classification make it unclear how to protect deep regression models. Third, most existing methods focus only on one stage of the copyright verification process and do not consider the holistic process. In this paper, we propose a three-stage holistic copyright verification method for deep regression models. In the first stage (defensive design stage), our method embeds watermarking by incorporating dropout operations with uncommon dropout rates into a deep regression model. In the second stage (covert detection stage), we propose a Bayesian network-based likelihood estimation method to infer the probability of copyright infringement based on the behavior of the suspicious model. If the probability is high, the third stage (overt response stage) follows, where the victim requests detailed information about the suspicious model. At this point, the KL divergence between the output distributions of the suspicious model and the victim model can be calculated to help determine whether infringement has occurred. We evaluate the effectiveness of our method on both an NLP task and a CV task, assessing four aspects including accuracy, robustness, fidelity, and efficiency. The experimental results demonstrate the advantages of our method. This study contributes to the field of AI governance by proposing a novel method for protecting the copyright of AI models.

Fields of Study & Subjects

Law AI and Intellectual Property Adversarial Robustness in Machine Learning Advanced Steganography and Watermarking Techniques

Publication details

Journal / source
Fundamental Research
Publication date
1 March 2026
Language
EN