Device Identification Using Relative Similarity Thresholds
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing device identification techniques face challenges in accurately determining new type devices using machine learning due to absolute similarity calculations and the difficulty in setting an appropriate new type determination threshold, especially in varying network environments.
Innovation Solution
A device identification apparatus that collects communication information, generates feature amounts, and uses machine learning to classify data with labels and a dummy label, calculating degrees of similarity and automatically setting a new type determination threshold based on the number of feature amounts classified into each label.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If machine learning is applied to calculate degree of similarity between identification target and accumulated data, then classification accuracy is improved, but the ability to determine new type devices is deteriorated due to relative amount calculation
Solution Approach 1:
The patent introduces a conversion mechanism that transforms machine learning's relative probability outputs into absolute similarity scores. This intermediary conversion layer allows the system to maintain the classification benefits of machine learning while restoring the ability to perform new type determination by comparing against thresholds, thus resolving the contradiction between improved classification accuracy and deteriorated new type detection capability.
2Measurement precision
If new type determination threshold is set high, then new type device detection accuracy is improved, but device identification reliability is deteriorated due to communication information variation
Solution Approach 1:
The patent implements dynamic threshold adjustment based on the distribution characteristics of similarity scores. Instead of using a fixed high threshold that causes misidentification, the system adapts the threshold level according to the actual data distribution and communication information variations, thereby maintaining high new type detection accuracy while preserving device identification reliability.
3Ease of operation
If simple machine learning application is used, then implementation ease is improved, but new type determination capability is deteriorated due to absolute amount requirement
Solution Approach 1:
The patent enables the machine learning system to self-serve the new type determination function by automatically converting its probability outputs into similarity scores that can be compared against thresholds. This self-service mechanism allows simple machine learning applications to maintain new type determination capability without requiring complex additional systems, thus resolving the contradiction between implementation ease and determination capability.
Data Source
AI summary
A device identification apparatus includes: a communication information collection unit configured to acquire communication information of existing devices and an identification target device; a feature amount generation unit configured to make the communication information of the existing devices' feature amounts and assign labels to generate first training data, make the communication information of the identification target device feature amounts and assign a dummy label to generate second training data, and further acquire communication information of the identification target device to generate identification data; a machine learning unit configured to cause a learning engine to learn the training data, and input the identification data to classify the identification data into the labels; a degree-of-similarity calculation unit configured to calculate a degree of similarity for each label; and a device identification unit configured to use a new type determination threshold to determine whether the device is a new type device.