# US 2019/0294955 A1 · Information processing apparatus, method, and computer readable storage medium

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2019/0294955 A1
- Document ID: US20190294955A1
- Status: Published application
- Kind: A1
- Application number: 16/424,840
- Filed: 2019-05-29
- Published: 2019-09-26
- Earliest claimed: 2017-01-12
- Assignees: Kddi Corporation
- Applicants: Kddi Corporation
- Inventors: Yusuke UCHIDA; Shigeyuki SAKAZAWA
- Web page: https://patentgrep.com/patent/US20190294955A1

## CPC

- G06N 3/045: Combinations of networks
- G06F 21/16: Program or content traceability, e.g. by watermarking
- G06N 3/0464: Convolutional networks [CNN, ConvNet]
- G06N 3/084: Backpropagation, e.g. using gradient descent
- G06N 3/09: Supervised learning
- G06N 20/20: Ensemble learning
- G06T 1/0021: {Image watermarking}
- G06N 3/048: Activation functions
- G06T 2201/0065: Extraction of an embedded watermark; Reliable detection
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06T: IMAGE DATA PROCESSING OR GENERATION, IN GENERAL

## Abstract

An information processing method for embedding watermark bits into weights of a first neural network includes: obtaining an output of a second neural network by inputting a plurality of input values obtained from a plurality of weights of the first neural network to the second neural network; obtaining second gradients of the respective plurality of input values based on an error between the output of the second neural network and the watermark bits; and updating the weights based on values obtained by adding first gradients of the weights of the first neural network that have been obtained based on backpropagation and the respective second gradients.

## Claims

### Claim 1

An information processing method for embedding watermark bits into weights of a first neural network, the method comprising:
obtaining an output of a second neural network by inputting a plurality of input values obtained from a plurality of weights of the first neural network to the second neural network;
obtaining second gradients of the respective plurality of input values based on an error between the output of the second neural network and the watermark bits; and
updating the weights based on values obtained by adding first gradients of the weights of the first neural network that have been obtained based on backpropagation and the respective second gradients.

### Claim 2 (depends on 1)

The information processing method according to claim 1,
wherein the first neural network includes N weight filters including a plurality of weights used for convolution, and
the plurality of input values are each an average value of weights of the N weight filters at the same position.

### Claim 3 (depends on 2)

The information processing method according to claim 2,
wherein a weight at a first position of the weight filters of the first neural network is updated based on a value obtained by adding the first gradient with respect to the weight and the second gradient with respect to a first input value, of the plurality of input values, which is an average value of weights of the N weight filters at the first position.

### Claim 4 (depends on 1)

The information processing method according to claim 1,
wherein the second neural network outputs a result obtained by performing a predetermined computation on the plurality of input values.

### Claim 5 (depends on 4)

The information processing method according to claim 4,
wherein the second neural network selects a same number of input values such as that of the watermark bits from the plurality of input values, and calculate an output by inputting each of the selected input values to an activation function.

### Claim 6 (depends on 4)

The information processing method according to claim 4,
wherein the second neural network selects a same number of pairs of input values such as that of the watermark bits from the plurality of input values, and calculate an output by inputting each of differences between input values of the respective selected pairs to an activation function.

### Claim 7

An information processing apparatus for embedding watermark bits into weights of a first neural network using the first neural network and a second neural network, the information processing apparatus comprising:
a calculating unit configured to obtain a plurality of input values from a plurality of weights of the first neural network;
an output calculating unit configured to obtain an output of the second neural network by inputting the plurality of input values to the second neural network;
a gradient calculating unit configured to obtain second gradients of the respective plurality of input values based on an error between an output of the second neural network and the watermark bits; and
a training unit configured to train the first neural network,
wherein the training unit is further configured to update the weights based on values obtained by adding first gradients of the weights of the first neural network that have been obtained based on backpropagation and the respective second gradients.

### Claim 8

A computer readable storage medium storing a program, the program, upon being executed by one or more processors in a computer, causing the computer to execute:
obtaining an output of a second neural network by inputting a plurality of input values obtained from a plurality of weights of a first neural network to the second neural network;
obtaining second gradients of the respective plurality of input values based on an error between the output of the second neural network and watermark bits; and
updating the weights based on values obtained by adding first gradients of the weights of the first neural network that have been obtained based on backpropagation and the respective second gradients.
