Quick and personalized model watermarking method for verifying model ownership
By generating user ID and watermark image datasets, training sub-network models and replacing them into the target model, the problem of the lack of personalization and high efficiency of existing black box watermark methods is solved, and a personalized watermark method for quickly verifying model ownership is realized.
WO2025065850A1 Publication Date: 2025-04-03ZHEJIANG UNIV +1
5 Cites 2 Cited by
Patent Information
- Application Number
- PCT/CN2023/134815
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-09-27
- Filing Date
- 2023-11-28
- Publication Date
- 2025-04-03
AI Technical Summary
Technical Problem
The existing black box watermarking methods lack personalization and cannot accurately locate a user. They need to train the model from scratch, which is cost-effective and has an impact on the performance of the original task of the model.
Method used
Generate user IDs through a random number generation algorithm, generate unique watermark image datasets using the generative model, train a subnet model, and replace it with the target model to achieve rapid verification of model ownership.
Benefits of technology
A personalized watermarking method is implemented, which can quickly verify model ownership, reduce training costs and time, and has negligible impact on the performance of the original task of the model.
✦ Generated by Eureka AI based on patent content.
Abstract
A quick and personalized model watermarking method for verifying model ownership. The method comprises: using a generative model to generate a watermark picture data set, which only belongs to a certain user; then using the picture data set to train a small network model which belongs to the user; then replacing a certain sub-network in a target model with the small network; and finally, the target model, part of the network of which has been replaced, working normally for an original classification task, wherein the target model shows a specific abnormal behavior regarding the generated watermark picture data set. Instead of retraining a model to make the model learn a watermark data set, a small network model including watermark data set information is trained, and the small network model is used to replace part of a target model. Moreover, watermark pictures owned by a user are generated by an adversarial generative network performing sampling on the basis of a user ID, such that the watermark data set owned by each user is unique.
Need to check novelty before this filing date? Find Prior Art