Edge Security Portfolio for Distributed Model Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning security approaches, particularly in AI-based health monitoring, face challenges in detecting malicious actions at the source, leading to increased server load, delayed detection, and exposure of user data, especially in pregnancy-related health monitoring where timely and secure data management is crucial.
Innovation Solution
Implementing a user device-level security portfolio that shifts model security analytics to the edge, using edge sensors to monitor malicious data manipulation, applying differential privacy techniques, and federated architecture to maintain data security and integrity while minimizing backend data upload, thus reducing central server load and ensuring low-latency security.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If current server-level security monitoring is used, then model security can be maintained, but server load increases and detection delays occur
Solution Approach 1:
The patent divides security monitoring into two segments: client-side monitoring (using local models to detect malicious actions) and server-side monitoring (for aggregate analysis). This segmentation reduces server load by processing security analytics at the edge devices where data originates, while maintaining comprehensive security coverage through coordinated operation of both layers.
Solution Approach 2:
The patent introduces local security models as intermediary components between data sources and the central server. These local models perform initial security analysis and classification, acting as mediators that filter and preprocess security-relevant information before transmission to the server, thereby reducing the volume of data requiring server processing.
2Reliability
If data is stored on user devices to prevent decryption, then data security improves, but modeling complexity increases and data completeness decreases
Solution Approach 1:
The patent extracts security-critical functionality from the central server and places it on client devices through deployment of local security models. This extraction maintains data security by keeping sensitive information on user devices while simplifying the overall system architecture by eliminating the need for complex centralized encryption and decryption infrastructure.
3Reliability
If federated architecture is used to share data, then data privacy is maintained, but model security against reverse engineering decreases
Solution Approach 1:
The patent applies preliminary security measures by classifying data at the client level before any transmission to the server. Local models perform advance security review and classification of data points, identifying and isolating potentially harmful data before it leaves the device. This preliminary action prevents malicious data from entering the federated learning process, protecting against reverse engineering attempts.
Solution Approach 2:
The patent converts the potential harm of data sharing into a benefit by using local security models to detect and classify malicious patterns. The local models transform what would be vulnerable data into secured data streams, enabling safe federated learning while maintaining privacy. The system turns the risk of data exposure into an opportunity for enhanced security through continuous local monitoring and classification.
Data Source
AI summary
The present disclosure is for systems and methods for data and model security in AI-based modeling approaches. Security techniques are applied at the user device level on edge devices to evaluate data and/or locally trained models for malicious content. Malicious content is detected and can be prevented from influencing central model updates or retraining.