LLM PII Unlearning Through Targeted Catastrophic Forgetting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current large language models (LLMs) inadvertently retain sensitive personally identifiable information (PII) during training, posing ethical risks and compliance challenges, and existing methods to erase PII require extensive retraining or fine-tuning, which is computationally prohibitive.
Innovation Solution
A system and method using targeted catastrophic forgetting (TCF) with adversarial fine-tuning, generating synthetic PII data with maximally different correlations to interfere with real PII gradients, incrementally fine-tuning the model to erase specific PII associations while preserving language proficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If extensive retraining or fine-tuning is performed to erase PII, then PII removal effectiveness is improved, but computational cost and time consumption increase significantly
Solution Approach 1:
The patent segments the PII removal process into two distinct phases: (1) identification phase where PII associations are detected in the trained model, and (2) removal phase where only the identified PII associations are targeted for erasure. This segmentation allows selective processing rather than complete retraining, significantly reducing computational cost while maintaining effective PII removal.
Solution Approach 2:
The patent applies partial action by performing only the necessary portion of retraining needed to remove PII. Instead of complete model retraining, the system performs targeted fine-tuning on specific PII-related parameters and associations, using minimal computational resources to achieve the desired privacy protection outcome.
2Reliability
If extensive retraining or fine-tuning is performed to erase PII, then PII removal effectiveness is improved, but training time increases significantly
Solution Approach 1:
The patent divides the training process into identification and removal segments, allowing the system to skip unnecessary retraining steps and directly target PII associations for erasure, thereby reducing training time while maintaining removal effectiveness.
Solution Approach 2:
The patent performs preliminary identification of PII associations before the removal process. By pre-identifying which parameters and associations contain PII, the system can then efficiently target only those specific elements for erasure, avoiding time-consuming full-model retraining.
3Reliability
If aggressive PII erasure methods are applied, then PII removal effectiveness is improved, but language model performance deteriorates
Solution Approach 1:
The patent applies local quality by treating different model parameters differently: PII-related parameters are targeted for erasure while non-PII parameters are preserved. This selective approach ensures that language model performance is maintained by protecting the quality of non-PII parameters while removing only the problematic PII associations.
Solution Approach 2:
The patent creates a copy or representation of the model parameters to identify PII associations, then applies erasure only to the identified PII-related copies or parameter subsets. This allows the original model structure and non-PII parameters to remain intact, preserving language model performance while achieving effective PII removal.
4Reliability
If complete model retraining is performed to remove PII, then thorough PII erasure is achieved, but resource consumption and computational overhead increase
Solution Approach 1:
The patent segments the model parameters into PII-related and non-PII-related subsets, allowing retraining resources to be allocated only to the PII subset. This segmentation achieves thorough PII erasure while minimizing resource consumption by excluding non-PII parameters from the retraining process.
Solution Approach 2:
The patent performs partial retraining only on the portion of the model that contains PII associations, rather than retraining the entire model. This partial action achieves sufficient PII erasure while dramatically reducing the quantity of computational resources and data processing required.
Data Source
AI summary
A method for removing unauthorized information associations from an LLM that is pre-trained on training data comprising unauthorized data (UD), the method including receiving a list of one or more UD instance-UD association pairs between real UD instances and UD associations identified for the real UD instances in the training data, generating one or more synthetic UD instance-UD association pairs comprising a synthetic UD instance-UD association pair from each real UD instance-UD association pair of the one or more real UD instance-UD association pairs, the synthetic UD instance-UD association pair being configured to one of reduce or remove influence of the real UD instance-UD association pair from which the synthetic UD instance-UD association pair was generated on an output of the LLM, and generating a fine-tuned LLM by iteratively fine-tuning the LLM based upon the one or more synthetic UD instance-UD association pairs.