Federated Model Aggregation Using Similarity Screening Against Poisoning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Federated learning technologies face challenges in ensuring data privacy and model robustness, with existing privacy protection methods like differential privacy and homomorphic encryption being computationally complex and affecting model availability and accuracy, while robustness verification methods like Krum, Trim-Mean, and Median discard normal models, impacting model aggregation.
Innovation Solution
A federated learning method that involves an aggregation device computing model similarities between user-generated segmented models and a standard model, selecting normal models for partial aggregation, and generating global aggregated models, while using a blockchain to verify user equipment trust and protect data privacy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If differential privacy or homomorphic encryption is used for privacy protection, then data privacy is protected, but computation complexity increases and model availability and accuracy are affected
Solution Approach 1:
The model is segmented into multiple parts (normal segmented models and suspected poisoned segmented models) based on similarity verification against a standard model. This segmentation allows the system to process different model portions differently, applying privacy protection selectively rather than uniformly across all data, thereby reducing overall computational complexity while maintaining privacy protection for critical segments.
Solution Approach 2:
Different privacy protection measures are applied to different segments of the model based on their similarity to the standard model. Normal segmented models undergo standard aggregation, while suspected poisoned models undergo more rigorous verification or are excluded. This local differentiation optimizes the balance between privacy protection and computational efficiency by applying stronger measures only where necessary.
2Reliability
If Krum, Trim-Mean, or Median methods are used for robustness verification, then model robustness is improved, but normal models or normal model parameters are discarded, affecting model aggregation
Solution Approach 1:
Before performing model aggregation, the system performs a preliminary similarity verification action by comparing each received segmented model against a standard model. This preliminary action identifies and flags suspected poisoned models before they can contaminate the aggregation process, allowing normal models to be preserved and aggregated without loss of their valuable parameters.
Solution Approach 2:
The system implements a feedback mechanism where the standard model serves as a reference for verifying incoming segmented models. Based on the similarity feedback from this verification, the system determines whether to include or exclude each model segment from aggregation. This feedback-driven approach ensures that normal models are retained while poisoned models are filtered out, preventing unnecessary loss of useful model parameters.
3Reliability
If existing robustness verification methods are used, then model robustness is enhanced, but model aggregation accuracy decreases due to improper screening of models and parameters
Solution Approach 1:
The system changes the verification parameter from simple statistical measures (mean, median, trim-mean) to similarity-based verification against a standard model. This parameter change enables more accurate identification of poisoned models while preserving normal models, thereby maintaining both robustness and aggregation accuracy. The similarity metric provides a more nuanced assessment of model quality compared to traditional robustness verification methods.
Data Source
AI summary
An aggregation device receives a plurality of first segmented models sent by user equipments, and separately computes a model similarity corresponding to each first segmented model. The aggregation device generates partial aggregated models based on second segmented models, where the second segmented models are selected from the plurality of first segmented models based on model similarities. The aggregation device aggregates partial aggregated models corresponding to the user equipments to generate global aggregated models. In the foregoing process, the aggregation device selects, from the first segmented models based on the model similarities corresponding to the first segmented models that are in a non-plaintext state, the second segmented models that can be used for partial aggregation, to generate the partial aggregated models, and then generates the global aggregated models based on the partial aggregated models.