Federated Learning Model Management for Privacy-Preserving Fraud Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Transaction processing systems face challenges in detecting anomalies and preventing fraud due to privacy measures that prohibit the transmission of user device data for centralized machine learning model training, leading to increased loss and decreased user engagement.
Innovation Solution
Implement a hybrid approach where machine learning model training is performed partially or fully on edge devices, with private user data remaining on the device, and utilize federated learning techniques to transmit obfuscated data to a central server for fine-tuning, while maintaining privacy and improving fraud detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If user device data is transmitted to central server for machine learning model training, then fraud detection capability is improved, but user privacy is compromised
Solution Approach 1:
The system segments the machine learning training process into two parts: (1) centralized model architecture design and initial training on the server, and (2) distributed fine-tuning on user devices using local data. This segmentation allows privacy-preserving training while maintaining fraud detection capability.
Solution Approach 2:
The patent introduces an intermediary mechanism where only model parameters and gradients (not raw user data) are transmitted between devices and server. This intermediary approach enables collaborative training while preserving user privacy by keeping sensitive data localized.
2Measurement precision
If centralized machine learning model training is performed, then fraud detection accuracy is improved, but data transmission latency increases
Solution Approach 1:
The system performs preliminary centralized training to establish a baseline model architecture and initial parameters before distributing it to user devices. This preliminary action reduces the need for repeated large-scale data transmissions, thereby reducing latency while maintaining detection accuracy.
Solution Approach 2:
The patent shifts the training dimension from centralized server-based to distributed device-based fine-tuning. By moving the training process to the edge dimension (user devices), the system reduces data transmission requirements and latency while maintaining model accuracy through federated learning techniques.
3Loss of information
If federated learning is implemented with local model training, then user privacy is maintained, but model performance consistency across devices decreases
Solution Approach 1:
The system implements a feedback mechanism where models trained locally on devices periodically upload performance metrics and updated parameters to the central server. The server aggregates these updates and redistributes refined model versions, ensuring consistent performance across diverse devices while maintaining privacy.
Solution Approach 2:
The patent employs parameter changes through federated averaging, where model parameters from multiple devices are aggregated and averaged to create a unified updated model. This parameter aggregation approach maintains performance consistency across devices while allowing local adaptation to device-specific data characteristics.
Data Source
AI summary
Techniques are disclosed in which a computer system receives, from a plurality of user computing devices, a plurality of device-trained models and obfuscated sets of user data stored at the plurality of user computing devices, where the device-trained models are trained at respective ones of the plurality of user computing devices using respective sets of user data prior to obfuscation. In some embodiments, the server computer system determines similarity scores for the plurality of device-trained models, wherein the similarity scores are determined based on a performance of the device-trained models. In some embodiments, the server computer system identifies, based on the similarity scores, at least one of the plurality of device-trained models as a low-performance model. In some embodiments, the server computer system transmits, to the user computing device corresponding to the low-performance model, an updated model.