# US 12,118,119 B2 · Federated private adversarial training

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

- Number: US 12,118,119 B2
- Document ID: US12118119B2
- Status: Granted
- Kind: B2
- Application number: 17/110,369
- Filed: 2020-12-03
- Published: 2022-06-09
- Granted: 2024-10-15
- Earliest claimed: 2020-12-03
- Assignees: International Business Machines Corporation
- Applicants: International Business Machines Corporation
- Inventors: Ngoc Minh Tran; Mathieu Sinn; Stefano Braghin
- Web page: https://patentgrep.com/patent/US12118119B2

## CPC

- G06F 21/6245: {Protecting personal data, e.g. for financial or medical purposes}
- G06F 18/2148: {characterised by the process organisation or structure, e.g. boosting cascade}
- G06F 21/602: {Providing cryptographic facilities or services}
- G06N 3/0442: characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
- G06N 3/0464: Convolutional networks [CNN, ConvNet]
- G06N 3/063: using electronic means
- G06N 3/08: Learning methods
- G06N 3/09: Supervised learning
- G06N 3/094: Adversarial learning
- G06N 3/096: Transfer learning
- G06N 3/098: Distributed learning, e.g. federated learning
- G06N 20/00: Machine learning
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

One or more computer processors transmit a machine learning model and an associated loss function to a worker, wherein the worker isolates private data. The one or more computer processors receive a plurality of encrypted gradients computed utilizing the transmitted machine learning model, the associated loss function, and the isolated private data. The one or more computer processors generate a plurality of adversarial perturbations, wherein the plurality of adversarial perturbations includes true perturbations and false perturbations. The one or more computer processors obfuscate the generated plurality of adversarial perturbations. The one or more computer processors transmit the obfuscated adversarial perturbations to the worker. The one or more computer processors harden the machine learning model utilizing the transmitted obfuscated adversarial perturbations and the private data.

## Claims

### Claim 1

A computer-implemented method comprising:
transmitting, by one or more computer processors, a machine learning model and an associated loss function to a worker, wherein the worker isolates private data from one or more adversarial strategies;
computing, by one or more computer processors, a plurality of encrypted gradients on the worker utilizing the transmitted machine learning model and associated loss function with the incorporation of the isolated private data;
generating, by one or more computer processors, a plurality of adversarial perturbations from the received plurality of encrypted gradients, wherein the plurality of adversarial perturbations includes true perturbations, that comprise adversarial noise that does change one or more subsequent machine learning model results, and false perturbations that comprise random noise that does not change one or more subsequent machine learning model results;
transmitting, by one or more computer processors, the generated adversarial perturbations to the worker, comprising:
transmitting, by one or more computer processors, one or more false perturbations to the worker until satisfying a perturbation requirement, wherein the false perturbations obscure at least one adversarial strategy;
adjusting, by one or more computer processors, the perturbation requirement based on a difficulty of adversarial identification; and
retraining, by one or more computer processors, the machine learning model utilizing the transmitted adversarial perturbations and the private data, comprising:
responsive to the machine learning model not converging, continuously streaming, by one or more computer processors, the generated adversarial perturbations to the worker.

### Claim 2 (depends on 1)

The computer-implemented method of claim 1, wherein transmitting the generated adversarial perturbations to the worker, comprises:
responsive to a probability not exceeding a probability threshold, transmitting, by one or more computer processors, a true perturbation to the worker.

### Claim 3 (depends on 1)

The computer-implemented method of claim 1, further comprising:
adjusting, by one or more computer processors, the perturbation requirement based on a, system importance, computing constraints, and temporal constraints.

### Claim 4 (depends on 1)

The computer-implemented method of claim 1, wherein transmitting the generated adversarial perturbations further comprises:
encrypting, by one or more computer processors, the generated adversarial perturbations with a public key associated with the worker.

### Claim 5 (depends on 1)

The computer-implemented method of claim 1, further comprising:
deploying, by one or more computer processors, the retrained machine learning model as a service in a cloud environment.

### Claim 6

A computer program product comprising:
one or more computer readable storage media and program instructions stored on the one or more computer readable storage media, the stored program instructions comprising:
program instructions to transmit a machine learning model and an associated loss function to a worker, wherein the worker isolates private data from one or more adversarial strategies;
program instructions to compute a plurality of encrypted gradients on the worker utilizing the transmitted machine learning model and associated loss function with the incorporation of the isolated private data;
program instructions to generate a plurality of adversarial perturbations from the received plurality of encrypted gradients, wherein the plurality of adversarial perturbations includes true perturbations, that comprise adversarial noise that does change one or more subsequent machine learning model results, and false perturbations that comprise random noise that does not change one or more subsequent machine learning model results;
program instructions to transmit the generated adversarial perturbations to the worker, wherein the program instructions further comprise:
program instructions to transmit one or more false perturbations to the worker until satisfying a perturbation requirement, wherein the false perturbations obscure at least one adversarial strategy;
program instructions to adjust the perturbation requirement based on a difficulty of adversarial identification; and
program instructions to retraining the machine learning model utilizing the transmitted adversarial perturbations and the private data, comprising:
program instructions to, responsive to the machine learning model not converging, continuously stream the generated adversarial perturbations to the worker.

### Claim 7 (depends on 6)

The computer program product of claim 6, wherein the program instructions to transmit the generated adversarial perturbations to the worker, comprises:
program instructions to, responsive to a probability not exceeding a probability threshold, transmit a true perturbation to the worker.

### Claim 8 (depends on 6)

The computer program product of claim 6, wherein the program instructions, stored on the one or more computer readable storage media, further comprise:
program instructions to adjust the perturbation requirement based on a difficulty of, system importance, computing constraints, and temporal constraints.

### Claim 9 (depends on 6)

The computer program product of claim 6, wherein the program instructions, stored on the one or more computer readable storage media, further comprise:
program instructions to encrypt the generated adversarial perturbations with a public key associated with the worker.

### Claim 10 (depends on 6)

The computer program product of claim 6, wherein the program instructions, stored on the one or more computer readable storage media, further comprise:
program instructions to deploy the retrained machine learning model as a service in a cloud environment.

### Claim 11

A computer system comprising: one or more computer processors;
one or more computer readable storage media; and
program instructions stored on the computer readable storage media for execution by at least one of the one or more processors, the stored program instructions comprising:
program instructions to transmit a machine learning model and an associated loss function to a worker, wherein the worker isolates private data from one or more adversarial strategies;
program instructions to compute a plurality of encrypted gradients on the worker utilizing the transmitted machine learning model and associated loss function with the incorporation of the isolated private data;
program instructions to generate a plurality of adversarial perturbations from the received plurality of encrypted gradients, wherein the plurality of adversarial perturbations includes true perturbations, that comprise adversarial noise that does change one or more subsequent machine learning model results, and false perturbations that comprise random noise that does not change one or more subsequent machine learning model results;
program instructions to transmit the generated adversarial perturbations to the worker, wherein the program instructions further comprise:
program instructions to transmit one or more false perturbations to the worker until satisfying a perturbation requirement, wherein the false perturbations obscure at least one adversarial strategy;
program instructions to adjust the perturbation requirement based on a difficulty of adversarial identification; and
program instructions to retraining the machine learning model utilizing the transmitted adversarial perturbations and the private data, comprising:
program instructions to, responsive to the machine learning model not converging, continuously stream the generated adversarial perturbations to the worker.

### Claim 12 (depends on 11)

The computer system of claim 11, wherein the program instructions to transmit the generated adversarial perturbations to the worker, comprises:
program instructions to, responsive to a probability not exceeding a probability threshold, transmit a true perturbation to the worker.

### Claim 13 (depends on 11)

The computer system of claim 11, wherein the program instructions, stored on the one or more computer readable storage media, further comprise:
program instructions to encrypt the generated adversarial perturbations with a public key associated with the worker.

### Claim 14 (depends on 11)

The computer system of claim 11, wherein the program instructions, stored on the one or more computer readable storage media, further comprise:
program instructions to deploy the retrained machine learning model as a service in a cloud environment.
