Generative Model Identity Unlearning via Latent Vector Replacement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing generative models face challenges in unlearning personal information, particularly images of individuals, due to the difficulty in accessing the entire learning dataset and high computational requirements of current unlearning methods.
Innovation Solution
A method for unlearning the identity of a person in a generative model using a computing device, involving the extraction of a source latent vector from a face image, setting a target latent vector to differ in identity, and applying local, adjacency-aware, and global preservation losses to remove the identity while maintaining model performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If access to the entire learning dataset is required for unlearning, then the completeness of unlearning is improved, but the difficulty of obtaining data and computational cost increase
Solution Approach 1:
The patent extracts only the essential information needed for unlearning - specifically the source latent vector from a single source image - rather than requiring access to the entire learning dataset. This extraction approach enables effective unlearning while avoiding the difficulty of obtaining and processing complete training data.
Solution Approach 2:
The unlearning process is segmented into distinct components: encoding the source image to obtain a source latent vector, identifying the specific identity information to be removed, and applying targeted modifications. This segmentation allows unlearning to be performed on a single image rather than requiring the entire dataset.
2Reliability
If access to the entire learning dataset is required for unlearning, then the completeness of unlearning is improved, but the computational cost increases
Solution Approach 1:
The patent extracts only the essential information needed for unlearning - specifically the source latent vector from a single source image - rather than requiring access to the entire learning dataset. This extraction approach enables effective unlearning while avoiding the difficulty of obtaining and processing complete training data.
Solution Approach 2:
Instead of performing unlearning on the entire dataset (excessive action), the patent applies unlearning selectively to a single source image and its corresponding latent vector (partial action). This partial approach achieves the necessary unlearning effect with significantly reduced computational resources.
3Object-affected harmful factors
If unlearning is performed to remove personal information, then privacy protection is improved, but the loss of useful information occurs
Solution Approach 1:
The patent applies local quality by targeting only the specific identity information in the latent space for modification, while preserving other useful information. The unlearning operation is localized to the identity attributes rather than affecting the entire dataset or model parameters globally, thus protecting privacy while retaining useful knowledge.
Solution Approach 2:
The latent space serves as an intermediary representation that allows selective manipulation of identity information without directly modifying the original training data or model weights. By operating in this intermediate latent space, the system can remove personal information while preserving the underlying generative capabilities and useful patterns.
Data Source
AI summary
A method of performing unlearning of people in a generative model includes inputting a source image including a face of a person to be unlearned in a pre-learned generative model into an encoder to extract a source latent vector in a latent space, setting a target latent vector so that the identity is different from that of a person corresponding to the source latent vector in the latent space, and performing unlearning to remove the identity of the person in the pre-learned model based on the source latent vector and the target latent vector.