Systems and methods for decentralized attribution of generative models

Patent metadata

Field	Value
Publication number	US-12518203-B2
Application number	US-202117544201-A
Country	US
Kind code	B2
Filing date	Dec 7, 2021
Priority date	Dec 7, 2020
Publication date	Jan 6, 2026
Grant date	Jan 6, 2026

Abstract

A system and associated methods for decentralized attribution of GAN models is disclosed. Given a group of models derived from the same dataset and published by different users, attributability is achieved when a public verification service associated with each model (a linear classifier) returns positive only for outputs of that model. Each model is parameterized by keys distributed by a registry. The keys are computed from first-order sufficient conditions for decentralized attribution. The keys are orthogonal or opposite to each other and belong to a subspace dependent on the data distribution and the architecture of the generative model.

First claim

What is claimed is:

1 . A system for decentralized attribution of generative models, comprising: a processor in communication with a memory, the memory including instructions executable by the processor to:
generate a key for attributing an output image of a first generative adversarial network (GAN) with a user device, wherein the key is one of a plurality of keys derived from first-order sufficient conditions for decentralized attribution of the key to the first GAN of the user device and distinguishability of the output image from an authentic image; transmit, to the user device and over a network, a set of application installation instructions executable at the user device to:
install an application including a GAN at the user device;
modify one or more model parameters of the GAN at the user device using the key to produce the first GAN uniquely installed at the user device; and
removing the key from the user device; and
apply a classification model to the output image of the first GAN generated at the user device, the output image of the first GAN being influenced by the key and the classification model being trained to detect the output image as being produced by the first GAN associated with the user device by correlating one or more features of the output image with the key.

2 . The system of claim 1 , wherein the plurality of keys are orthogonal keys that achieve distinguishability and attributability.

3 . The system of claim 1 , wherein the plurality of keys belong to a subspace dependent upon data distribution and architecture of each GAN.

4 . The system of claim 1 , wherein the first GAN is one of a plurality of GANs distributed to a plurality of end user devices the plurality of GANS being respectively and uniquely modified using the plurality of keys.

5 . A method for decentralized attribution of generative models, comprising:
accessing, by a processor, information associated with a dataset, the dataset configured for GAN model generation for a plurality of end user devices; generating, by a first processor in communication with a memory, a key of a plurality of keys for perturbation of one or more model parameters of a GAN model, each key of the plurality of keys being computed from first-order sufficient conditions for decentralized attribution of respective outputs of a plurality of perturbed GAN models with an associated key of the plurality of keys; verifying, by the processor, attribution of the key to a perturbed GAN model of the plurality of perturbed GAN models by implementing a linear classifier that returns positive only for outputs of the perturbed GAN model; transmitting, from the processor and over a network, the GAN model and the key to an end user device of the plurality of end user devices; perturbing, upon installation of the GAN model at a memory of the end user device, one or more model parameters of the GAN model based on the key to produce the perturbed GAN model such that an output generated by the perturbed GAN model at the user device is influenced by the key; and detecting, by application of the output of the GAN model as input to a classifier model, the output of the GAN model as being produced by the end user device by correlating one or more features of the output with the key.

6 . The method of claim 5 , the GAN model being one of a plurality of GAN models respectively distributed to the plurality of end user devices, the plurality of GAN models being respectively and uniquely modified using the plurality of keys upon installation at the plurality of end user devices.

7 . The method of claim 5 , further comprising installing an application to the end user device, the application including the GAN model and the key, the installation modifying the GAN model according to the key to output the perturbed GAN model verifiable by the key, and subsequently removing the key from the end user device.

Assignees

• Yang Yezhou
• Kim Changhoon
• Ren Yi
• Univ Arizona State

Inventors

• Yang Yezhou
• Kim Changhoon
• Ren Yi

Classifications

• G06F21/64

  Protecting data integrity, e.g. using checksums, certificates or signatures · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/0475

  Generative networks · CPC title

• G06N3/094

  Adversarial learning · CPC title

• G06N3/045

  Combinations of networks · CPC title