ML Firewall Filtering for Invalid Input and Output Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing technologies fail to prevent the input of invalid data, such as malicious or error data, to machine-learned information processing devices, which can compromise the security and integrity of machine learning models.
Innovation Solution
An information processing device functioning as a firewall is provided, equipped with input and output data check units that include analysis and filtering units to validate and filter out invalid data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional authorization control is implemented for machine learning models, then model access security is improved, but the system cannot handle invalid data including malicious attacks, user errors, and outdated training data
Solution Approach 1:
The system divides data validation into separate functional modules: input data analysis unit, input data filtering unit, output data analysis unit, and output data filtering unit. Each unit handles specific validation tasks independently, allowing comprehensive data checking without monolithic complexity.
Solution Approach 2:
The patent introduces a firewall as an intermediary component between external data sources and the machine learning server. This firewall acts as a mediator that analyzes and filters both input and output data, preventing direct interaction between potentially harmful external data and the vulnerable ML model.
2Reliability
If data filtering is performed to remove invalid data, then system security is improved, but processing time and computational resources increase
Solution Approach 1:
The system performs data analysis and filtering in advance before data reaches the machine learning model. The input data analysis and filtering units pre-validate all incoming data, and the output data analysis and filtering units prepare responses beforehand, preventing security issues rather than reacting to them.
Solution Approach 2:
The patent extracts and removes harmful components from data streams through dedicated filtering units. The input data filtering unit extracts malicious or invalid input data before it can affect the model, while the output data filtering unit extracts potentially harmful model responses before they reach external systems.
Data Source
AI summary
Invalid data is prevented from being input to a machine-learned information processing device from outside. An information processing device functioning as a firewall for a machine learning server is provided. The information processing device includes an input data check unit and an output data check unit. The input data check unit includes an input data analysis unit and an input data filtering unit. The output data check unit includes an output data analysis unit and an output data filtering unit. The input data analysis unit is configured to analyze input data to be input to the machine learning server. The input data filtering unit is configured to filter out less valid data from the input data. The output data analysis unit is configured to analyze output data outputted from the machine learning server. The output data filtering unit is configured to filter out less valid data from the output data.