Federated Model Clustering for Malicious Update Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for federated learning in fraudulent transaction detection face challenges due to skewed data distributions among banks and malicious model updates, leading to robustness issues and false positives, with no holistic solution addressing both problems simultaneously.
Innovation Solution
A method and system for federated learning that involves preprocessing local classification models, clustering them using techniques like PCA and k-means, selecting active models based on epsilon cluster selection and Silhouette coefficients, and randomly choosing a global model to enhance robustness and detect malicious clients.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional federated learning methods are used to train models across multiple banks, then the model can learn from diverse fraudulent transaction patterns, but the skewed data distribution among banks degrades model robustness
Solution Approach 1:
The patent segments the federated learning process into distinct phases: initial centralized training to establish a baseline model, followed by federated learning rounds where local models are trained and evaluated. This segmentation allows the system to handle data heterogeneity by processing banks with sufficient data separately from those with limited data, thereby maintaining robustness while achieving adaptability.
Solution Approach 2:
The patent dynamically adjusts training parameters including learning rates, batch sizes, and evaluation thresholds based on each bank's data characteristics. By changing these parameters adaptively rather than using fixed values, the system maintains model robustness across banks with skewed data distributions while preserving the ability to learn diverse fraudulent patterns.
2Adaptability or versatility
If all local models are aggregated in federated learning, then comprehensive fraud detection coverage is achieved, but malicious model updates from compromised clients introduce false positives
Solution Approach 1:
The patent introduces an intermediary evaluation phase between model aggregation and deployment. Local models are first evaluated against a ground truth dataset before being aggregated into the global model. This intermediary step acts as a filter that prevents malicious or poorly performing models from contaminating the global model, thereby reducing false positives while maintaining comprehensive fraud detection coverage.
Solution Approach 2:
The patent implements a feedback mechanism where model performance is continuously evaluated against ground truth data, and this evaluation information is used to adjust aggregation weights and filtering criteria. The feedback loop enables the system to identify and mitigate the impact of malicious updates while preserving beneficial contributions from legitimate clients, thus maintaining both coverage and reliability.
3Measurement precision
If data is centralized for training, then model accuracy improves through access to all transaction patterns, but data privacy and security requirements prevent actual centralization
Solution Approach 1:
The patent introduces a ground truth dataset as an intermediary that enables accurate model evaluation without requiring centralization of actual transaction data. The ground truth serves as a reference standard that allows the system to achieve high model accuracy through federated learning while maintaining data privacy, as no sensitive transaction data needs to be centralized.
Data Source
AI summary
The present disclosure provides a federated learning based identification of non-malicious classification models where the conventional model fails to perform. Initially, the system receives a local classification model from each of a plurality of clients. Further, a set of one-dimensional arrays are obtained based on a plurality of local classification models associated with the plurality of clients using a flattening technique. Further, a major cluster and a minor cluster are obtained by clustering the set of one-dimensional arrays using a clustering technique. After clustering, a plurality of active classification models are selected based on the major cluster and the minor cluster using an epsilon cluster selection technique. Further, a global classification model is selected from the plurality of active models using a random selection technique. Finally, the selected global classification model is transmitted to each of the plurality of clients.