Federated Fine-Tuning With Utility-Aware Noise Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge of balancing model availability and data privacy in the context of model-as-a-service (MaaS) ecosystems, where pre-trained language models are fine-tuned using client datasets, is exacerbated by the risk of model and data privacy leakage.
Innovation Solution
A model federated fine-tuning method that involves segmenting text samples, identifying key words with classification utility, applying noise perturbation only to non-utility words, and collaboratively fine-tuning model parameters with a server and clients, ensuring privacy protection while maintaining model performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If noise perturbation is applied to all embedding vectors during federated fine-tuning, then data privacy is protected, but model performance deteriorates due to excessive noise on utility words
Solution Approach 1:
The patent applies different noise perturbation strategies to different parts of the embedding vectors based on their classification utility. Words are divided into utility words (important for classification) and non-utility words, with only non-utility words receiving noise perturbation. This local differentiation resolves the contradiction by protecting privacy through noise while preserving model performance on critical words.
Solution Approach 2:
The patent segments the text input into utility words and non-utility words based on classification importance. This segmentation allows differential privacy protection where utility words retain their original embedding vectors for accurate classification, while non-utility words receive noise perturbation for privacy protection, thus resolving the performance-privacy tradeoff.
2Adaptability or versatility
If all embedding vectors are transmitted to the server for fine-tuning, then model availability is improved, but data privacy risk increases
Solution Approach 1:
The patent extracts and identifies utility words from the text that are critical for classification tasks. These utility words have their embedding vectors protected from noise perturbation and transmitted to the server, while non-utility words are perturbed. This extraction approach enables model fine-tuning with protected data, balancing availability and privacy.
Solution Approach 2:
The patent introduces an intermediary mechanism (utility word identification and differential noise application) between the raw embedding vectors and the server transmission. This intermediary selectively applies noise based on word utility, acting as a mediator that preserves both privacy and model performance simultaneously.
3Adaptability or versatility
If pre-trained models are fine-tuned on client datasets in MaaS ecosystem, then model customization is achieved, but model and data privacy leakage risks arise
Solution Approach 1:
The patent applies local quality differentiation by treating utility words and non-utility words differently during the fine-tuning process. Utility words essential for model customization are preserved without noise, while non-utility words are perturbed for privacy protection. This enables customized model training while maintaining privacy security.
Solution Approach 2:
The patent applies partial noise perturbation only to non-utility words rather than all words. This partial action approach provides sufficient privacy protection for non-critical information while maintaining the integrity of utility words needed for model customization, achieving the right balance between customization and security.
Data Source
AI summary
A model federated fine-tuning method, a text classification method and apparatus, a medium, and a device. The method includes: performing word segmentation on a text sample, and generating an embedding vector corresponding to each of a plurality of segmented words by using a second model; determining, from the plurality of segmented words, a target segmented word having classification utility for a text category marked by a classification label; performing noise perturbation processing on embedding vectors corresponding to other segmented words in the plurality of segmented words except the target segmented word to obtain a perturbation vector; and collaboratively fine-tuning a model parameter of a first model with a server and another client based on the perturbation vector, an embedding vector corresponding to the target segmented word, and the classification label.