System and method for AI model watermarking

Patent metadata

Field	Value
Publication number	US-12417394-B2
Application number	US-202117204935-A
Country	US
Kind code	B2
Filing date	Mar 17, 2021
Priority date	Mar 17, 2021
Publication date	Sep 16, 2025
Grant date	Sep 16, 2025

Abstract

Method and system for watermarking prediction outputs generated by a first AI model to enable detection of a target AI model that has been distilled from the prediction outputs. Includes receiving, at the first AI model, a set of input data samples from a requesting device; storing at least a subset of the input data samples to maintain a record of the input data samples; predicting, using the first AI model, a respective set of prediction outputs that each include a probability value, the AI model using a watermark function to insert a periodic watermark signal in the probability values of the prediction outputs; and outputting, from the first AI model, the prediction outputs including the periodic watermark signal.

First claim

The invention claimed is:

1. A method for watermarking prediction outputs generated by a first AI model to enable detection of a target AI model that has been distilled from the watermarked prediction outputs, comprising:
receiving, at the first AI model, a set of input data samples from a requesting device;
storing at least a subset of the input data samples to maintain a record of the input data samples;
predicting, using the first AI model, a respective set of watermarked prediction outputs that each include a probability value, the first AI model using a watermark embedding function to insert a periodic watermark signal in the probability values to generate the watermarked prediction outputs; and
outputting, from the first AI model, the watermarked prediction outputs including the periodic watermark signal.

2. The method of claim 1 wherein the periodic watermark signal is configured to cause an AI model that is distilled from the respective set of prediction outputs to insert a periodic signal into prediction outputs of the AI model that can be detected as matching the periodic watermark signal.

3. The method of claim 1 , further comprising, prior to the predicting, including the watermark embedding function in a preliminary AI model to generate the first AI model.

4. The method of claim 3 , wherein the preliminary AI model is an untrained model, the method comprising training the preliminary AI model using a loss that is based on the outputs of the preliminary AI model with the included watermark embedding function.

5. The method of claim 1 , comprising defining a key corresponding to the watermark embedding function, the key including a random projection vector, wherein the watermark embedding function inserts the periodic watermark signal based on the random projection vector.

6. The method of claim 5 , further comprising determining if the target AI model has been distilled from the first AI model by:
submitting a query to the target AI model that includes at least some of the stored subset of the input data samples;
receiving prediction outputs from the target AI model corresponding to the input data samples;
determining, based on the key, if a periodic signal that matches the periodic watermark signal can be detected in the prediction outputs from the target AI model.

7. The method of claim 6 wherein the key further includes information that identifies a frequency of the periodic watermark signal and a target prediction output to monitor for the periodic watermark signal.

8. The method of claim 1 , wherein the watermark embedding function is configured to modify softmax outputs of a softmax layer of the first AI model by inserting the periodic watermark signal into the probability values of the prediction outputs.

9. A method of determining if a target AI model has been distilled from a first AI model by:
submitting a query to the target AI model that includes input data samples that were previously provided to the first AI model;
receiving prediction outputs from the target AI model corresponding to the input data samples; and
determining, based on a predetermined key, if a periodic signal that matches a known periodic watermark signal can be detected in the prediction outputs from the target AI model.

10. The method of claim 9 wherein the determining comprises:
determining, based on a Fourier power spectrum of the prediction outputs and a projection vector included in the predetermined key, if a signal power that corresponds to the frequency of the known periodic watermark signal can be detected in the prediction outputs from the target AI model.

11. A computer system comprising one or more processing units and one or more non-transient memories storing computer implementable instructions for execution by the one or more processing devices, wherein execution of the computer implementable instructions configures the computer system to perform a method for watermarking prediction outputs generated by a first AI model to enable detection of a target AI model that has been distilled from the watermarked prediction outputs, comprising:
receiving, at the first AI model, a set of input data samples from a requesting device;
storing at least a subset of the input data samples to maintain a record of the input data samples;
predicting, using the first AI model, a respective set of watermarked prediction outputs that each include a probability value, the first AI model using a watermark embedding function to insert a periodic watermark signal in the probability values to generate the watermarked prediction outputs; and
outputting, from the first AI model, the watermarked prediction outputs including the periodic watermark signal.

12. The computer system of claim 11 wherein the periodic watermark signal is configured to cause an AI model that is distilled from the respective set of prediction outputs to insert a periodic signal into prediction outputs of the AI model that can be detected as matching the periodic watermark signal.

13. The computer system of claim 12 wherein the method includes, prior to the predicting, including the watermark embedding function in a preliminary AI model to generate the first AI model.

14. The computer system of claim 13 wherein the preliminary AI model is an untrained model, the method comprising training the preliminary AI model using a loss that is based on the outputs of the preliminary AI model with the included watermark embedding function.

15. The computer system of claim 13 , wherein the method includes defining a key corresponding to the watermark embedding function, the key including a random projection vector, wherein the watermark embedding function inserts the periodic watermark signal based on the random projection vector.

Assignees

• Huawei Cloud Computing Tech Co Ltd

Inventors

• Charette Laurent
• Chu Lingyang
• Wang Lanjun
• Zhang Yong

Classifications

• G06N20/00

  Machine learning · CPC title

• G06N3/04

  Architecture, e.g. interconnection topology · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/0499

  Feedforward networks · CPC title

• G06F21/16

  Program or content traceability, e.g. by watermarking · CPC title