Behavioral Authentication Using CNN Feature Extraction

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Conventional user authentication systems using behavioral data from multiple sensors on mobile devices face inaccuracies due to unaligned and varying sample lengths, relying on domain knowledge for feature extraction, which limits their accuracy and compatibility with machine learning architectures.

Innovation Solution

The method normalizes samples from multiple sensors to have a consistent length, allowing a convolutional neural network to extract features for user authentication, thereby improving accuracy and compatibility with machine learning architectures.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If unaligned samples with different lengths are used from multiple sensors, then the system can capture diverse behavioral data, but the data is incompatible with machine learning architecture and reduces authentication accuracy

Engineering Contradiction:

Improvebehavioral data capture capabilityVSAvoidauthentication accuracy

Core Design Contradiction:

Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

The patent applies preliminary action by normalizing sensor samples to a predetermined fixed length before inputting them to the neural network. This preprocessing step ensures that all behavioral data from multiple sensors with different sampling rates and durations are standardized in advance, making them compatible with the machine learning architecture and enabling accurate authentication

Inventive Principle:

Principle #10Preliminary action

2Ease of manufacture

If domain knowledge is used to extract features from behavioral data, then the system can process the data, but it relies on expert quality and reduces authentication accuracy

Engineering Contradiction:

Improvefeature extraction processabilityVSAvoidauthentication accuracy

Core Design Contradiction:

Ease of manufactureVSMeasurement precision

Solution Approach 1:

The patent replaces the manual domain knowledge-based feature extraction process with an automated convolutional neural network. The CNN automatically learns and extracts relevant features from the normalized behavioral data without requiring expert intervention, thereby eliminating reliance on subjective expert quality and improving authentication accuracy through data-driven feature extraction

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

3Quantity of substance

If samples from multiple sensors are used, then the system can gather comprehensive behavioral data, but the varying sample lengths create alignment issues

Engineering Contradiction:

Improvebehavioral data volumeVSAvoiddata alignment complexity

Core Design Contradiction:

Quantity of substanceVSDevice complexity

Solution Approach 1:

The patent changes the parameter of sample length from variable to fixed by normalizing all sensor outputs to a predetermined length. This parameter transformation resolves the alignment complexity by ensuring that temporal data from sensors with different sampling rates and durations can be uniformly processed, maintaining comprehensive behavioral data coverage while simplifying the input structure for the neural network

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentUS10061909B2Device authentication based on behavior classification using convolution neural network

Publication Date: 2018.08.28 QUALCOMM INC

AI summary

A method of authenticating a user on a mobile device includes gathering samples of behavioral data of the user from multiple sensors of the mobile device, each sensor generating a different number of samples. The method also includes normalizing the samples to have a same number of samples for each sensor. The method further includes extracting, with a convolutional neural network, features from the normalized samples and authenticating the user based on the extracted features.