Machine Unlearning in Generative Models via Loss Divergence

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Generative models face challenges in unlearning regulated or sensitive data, as existing methods are focused on classification tasks and impractical for image-to-image models, and often lack access to the original training dataset, posing risks for user privacy and copyright compliance.

Innovation Solution

A machine unlearning process that involves a target model and an original model, using a forget dataset and a retain dataset, where the process iteratively computes losses based on inputs from Gaussian noise and weighted combinations to minimize the expectation value, effectively 'forgetting' sensitive data while preserving model performance.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Manufacturing precision

If generative models are trained on comprehensive datasets to improve content generation quality, then model performance is improved, but user privacy and copyright compliance risks increase

Engineering Contradiction:

Improvecontent generation qualityVSAvoidprivacy and copyright risks

Core Design Contradiction:

Manufacturing precisionVSObject-affected harmful factors

Solution Approach 1:

The patent extracts and removes specific sensitive data patterns from the model's learned knowledge by computing a divergence objective between the original model and a modified model that has forgotten the sensitive data. This allows the model to retain general generation capabilities while removing specific harmful knowledge about regulated content.

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The patent changes the model's parameters through iterative optimization of a loss function that balances maintaining generation quality on normal data versus forgetting sensitive data patterns. By adjusting the optimization objective and iterating, the model parameters are modified to achieve both goals simultaneously.

Inventive Principle:

Principle #35Parameter changes

2Object-affected harmful factors

If models are retrained from scratch with filtered datasets to mitigate risks, then privacy and copyright compliance is improved, but time and computational resources increase significantly

Engineering Contradiction:

Improveprivacy and copyright complianceVSAvoidretraining time

Core Design Contradiction:

Object-affected harmful factorsVSLoss of time

Solution Approach 1:

The patent performs preliminary actions by computing divergence objectives and iterating on loss functions before full retraining would be required. This allows the model to be progressively modified to forget sensitive data without starting from scratch, significantly reducing the time and computational resources needed compared to complete retraining.

Inventive Principle:

Principle #10Preliminary action

3Object-affected harmful factors

If comprehensive filtering of regulated content from the original dataset is performed, then privacy and copyright compliance is improved, but the original dataset is often unavailable making this impractical

Engineering Contradiction:

Improveprivacy and copyright complianceVSAvoiddata filtering feasibility

Core Design Contradiction:

Object-affected harmful factorsVSEase of manufacture

Solution Approach 1:

The patent introduces an intermediary approach by working with a subset of available data (retain dataset) and using computational objectives to indirectly remove sensitive knowledge without needing access to the complete original training dataset. This mediator approach allows compliance to be achieved even when the original dataset is unavailable.

Inventive Principle:

Principle #24Intermediary (Mediator)

4Device complexity

If existing unlearning algorithms are applied to classification tasks only, then implementation complexity is reduced, but applicability to generative models is limited

Engineering Contradiction:

Improveunlearning algorithm complexityVSAvoidapplicability to generative models

Core Design Contradiction:

Device complexityVSAdaptability or versatility

Solution Approach 1:

The patent creates a universal unlearning framework that can be applied to both classification tasks and generative models by formulating a general divergence objective that works across different model types. This allows the same fundamental approach to be used for various AI systems, greatly enhancing adaptability while maintaining manageable complexity.

Inventive Principle:

Principle #6Universality (Multi-functionality)

Data Source

PatentUS20250103878A1Systems and methods for machine unlearning in generative models

Publication Date: 2025.03.27 JPMORGAN CHASE BANK NA

AI summary

In some aspects, the techniques described herein relate to a method including: providing a first datum to a target model, wherein the first datum is retrieved from a forget dataset; providing a sample drawn from Gaussian noise to an original model; computing a first loss, wherein the first loss is based on target model output from processing the first datum and original model output from processing the sample drawn from Gaussian noise; providing a second datum to the target model, wherein the second datum is retrieved from a retain dataset; providing the second datum to the original model as input to the original model; computing a second loss, wherein the second loss is based on target model output from processing the second datum and original model output from processing the second datum; and combining the first loss and the second loss with an alpha weighting to generate a weighted combination.