CN-122021882-A - Forgetting learning method based on causal knowledge decoupling

Abstract

The invention discloses a forgetting learning method based on causal knowledge decoupling, and aims to solve the balance problem that excessive forgetting (low fidelity) or incomplete forgetting (low effectiveness) easily occurs when the existing forgetting learning technology processes shared knowledge. The method is characterized in that the confounding factor of shared knowledge is identified through causal analysis, and is decoupled into attribute-level semantics for fine removal. The method mainly comprises the steps of firstly decomposing sample characteristics into attribute semantics shared across categories and unique attribute semantics specific to the categories by utilizing a knowledge decoupling module based on a variational self-encoder (beta-VAE), secondly utilizing a counterfactual reasoning generation module to reserve the unique attributes of the categories and the shared attributes of the categories to be forgotten through logic combination to synthesize counterfactual samples positioned near decision boundaries, thirdly utilizing the generated counterfactual samples to serve as anchor points to construct decision boundaries under a contrast learning framework by utilizing a knowledge refining module, and finally conducting constraint optimization by utilizing a reconstructed forgetting learning loss function, wherein the loss function consists of Multiple Reserved Loss (MRL) of ensuring the performance of the model in the non-target categories and Contrast Forgetting Loss (CFL) for enhancing the rigidity of the forgetting boundary, so that the balance of fidelity and effectiveness forgetting is realized. According to the invention, through decoupling and reconstruction of attribute-level semantics, robustness, generalization capability and calculation efficiency of the model after the data forgetting request is executed are obviously improved.

Inventors

• WU YIRUI
• XIA YUHANG

Assignees

• 河海大学

Dates

Publication Date

20260512

Application Date

20260113

Claims (1)

1. 1. The forgetting learning method based on causal knowledge decoupling is characterized by comprising the following steps of: 1. Knowledge decoupling module This module is responsible for performing feature extraction and decoupling operations based on a Structural Causal Model (SCM) aimed at identifying and separating confounding factors of shared knowledge. The module decouples the unobservable properties into mutually independent semantic groups by constructing an encoder-decoder structure: The semantic feature coding comprises the steps of extracting shared attribute semantics S existing across categories in a sample by using a first encoder Q φ , extracting unique attribute semantics U with category identification by using a second encoder Q ψ , constructing a multi-dimensional semantic, namely constructing an observation classifier theta O , ensuring that the shared attribute semantics S captures an observation context O irrelevant to the category by minimizing observation loss, constructing a category classification theta Y , ensuring that the unique attribute semantics U captures a core category label Y by minimizing category loss, and optimizing by using a weighted Evidence Lower Bound (ELBO) loss function, wherein the function combines reconstruction loss (used for restoring original sample characteristics X) and KL divergence items (used for restraining independence of latent variable distribution) regulated based on super-parameters beta, so that attribute decoupling is realized in a semantic space, and interference of a confounding factor is eliminated. 2. Knowledge refining module combined with anti-facts reasoning The module is responsible for carrying out imagination reasoning by using the decoupled attribute semantics, and realizing the balance performance between forgetting and reservation by constructing a counterfactual sample refining model decision boundary: The method comprises the steps of generating a trace-back sample, extracting a shared attribute characteristic direction from a sample to be forgotten, performing an intervention operation, namely randomly sampling a unique attribute characteristic vector from a reserved sample, performing an intervention action to combine the two, generating a prediction, namely, accurately simulating a decision blind area combining 'core semantics of the reserved category' and 'shared background of the forgotten category' in a characteristic space by generating a trace-back sample, namely, decoding the combined vector into the trace-back sample, carrying out gradient update on the target model by weighting total loss, carrying out forgetting optimization based on reconstruction loss, namely, multiple reservation constraint, namely, calculating KL divergence of a predictive probability distribution of the original model and the model to be updated on the trace-back sample, reserving the shared semantics irrelevant to a forgetting target by the constraint model, and preventing excessive forgetting, and comparing a forgetting mechanism, namely, constructing a comparison learning loss, defining the trace-back sample and the sample to be forgotten as a positive sample pair to pull a characteristic distance, and simultaneously defining the trace-back sample pair to the trace-back sample pair, so as to enhance the rigidity of a forgetting boundary, and carrying out gradient update on the target model by weighting total loss, and carrying out gradient update on the target model by the weighted total loss, so that the optimal performance balance between the Fidelity (Fidelity) and the effectiveness (EFFECTIVENESS) of the model after forgetting request is achieved.

Description

Forgetting learning method based on causal knowledge decoupling Technical Field The invention relates to the technical field of artificial intelligence and data privacy protection, in particular to a forgetting learning method for deep neural network privacy erasure, and especially relates to a forgetting learning method based on causal knowledge decoupling and anti-facts reasoning. Background With the development of data privacy protection awareness and related regulations (such as General Data Protection Regulations (GDPR)), users are given "forgotten rights". Forgetting learning (Machine Unlearning) has become a research hotspot as a technique to remove specific private data from a trained model without retraining from the head. Its core goal is to maintain the Fidelity of the model to the retained data (i.e., the ability to retain the remaining data) while ensuring validity (EFFECTIVENESS, i.e., the complete removal of the target data). However, the existing forgetting learning method still faces serious challenges in practical application. The invention is discovered through deep causal points, and the limitation of the prior art is mainly caused by improper treatment of shared knowledge: 1. Excessive forgetting, namely, because semantic overlapping (such as 'hairiness', 'quadruped' and other visual attributes shared by 'cat' and 'tiger') exists between a target class (data to be forgotten) and a reserved class, the existing method is extremely easy to delete the knowledge shared by the cross classes by mistake when the target class is removed, so that the performance of the model is greatly reduced when the reserved class data is processed, namely, the fidelity is low. 2. Partial approaches tend to preserve common features in order to maintain fidelity. However, these residual feature mappings create confounding effects in the feature space, so that the model can still implicitly identify category attributes that should be forgotten by sharing features, resulting in incomplete forgetting, i.e., low effectiveness. Interference of confounding factor-as is known by Structural Causal Model (SCM) analysis, the observed sample features contain "shared attributes" and "unique attributes". In existing forgetting learning processes, shared knowledge plays the role of confounding factors, which create a non-causal path between the forgetting category and the prediction result. Most of the prior art only carries out heuristic fine tuning at the model parameter level, and lacks a mechanism for effectively decoupling shared knowledge from unique knowledge at the semantic level. 4. Ambiguity of forgetting boundary because of lack of fine control on attribute-level semantics, the existing method is difficult to define an accurate forgetting decision boundary, and particularly when a category with high semantic similarity is processed, a rigid boundary which can thoroughly clear target privacy and accurately protect common features cannot be established. Therefore, how to identify and decouple shared knowledge from the causal layer and reconstruct a more rigid decision boundary by using decoupled semantic information to realize balance is a key technical problem to be solved urgently in the forgetting learning field at present. Disclosure of Invention Aiming at the technical problem that excessive forgetting (low fidelity) or incomplete forgetting (low effectiveness) is easy to occur when the existing forgetting learning technology processes shared knowledge with semantic overlapping, the invention provides a forgetting learning method based on causal knowledge decoupling. The method identifies and decouples the confounding factor of shared knowledge through causal analysis, thereby realizing high-efficiency forgetting of target data while ensuring that the model has high fidelity to reserved data. In order to achieve the above purpose, the invention adopts the following technical scheme: The invention provides a forgetting learning method based on causal knowledge decoupling, which has the core ideas that a Structural Causal Model (SCM) is utilized to identify and decouple a confounding factor of shared knowledge, the distribution of a decision boundary area is simulated through anti-facts reasoning, and the model is refined by utilizing a reconstructed loss function under a comparative learning framework. The method can effectively solve the conflict of excessive forgetting and incomplete forgetting in the traditional method, and can realize high-efficiency erasure of target data while ensuring that the model has high fidelity to reserved data, thereby achieving optimal balance performance. The method mainly comprises the following core modules and steps: 1. Knowledge decoupling module in combination with β -VAE: The module is responsible for executing feature extraction and decoupling operations based on causal logic, and decomposing complex sample features into mutually independent attribute-level semantic groups