Dual-Model Voice Authentication Against Impersonation Attacks

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing systems are vulnerable to voice impersonation attacks, which can lead to malicious activities within organizations, as they rely solely on acoustic features that can be easily mimicked, lacking robust authentication mechanisms.

Innovation Solution

A dual-machine learning approach combining acoustic and linguistic models to analyze voice signals, using acoustic models for phonic characteristics and grammatical models for linguistic patterns, generating an authenticity score to distinguish genuine users from impersonators.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of operation

If acoustic features are used for voice authentication, then the authentication process is simple and fast, but the system becomes vulnerable to voice impersonation attacks

Engineering Contradiction:

Improveauthentication process simplicityVSAvoidauthentication security

Core Design Contradiction:

Ease of operationVSReliability

Solution Approach 1:

The patent combines acoustic modeling (AM) and linguistic modeling (LM) into a unified authentication system. The acoustic model analyzes phonetic characteristics of voice signals, while the linguistic model analyzes linguistic patterns. Both models generate scores that are combined to produce a final authentication decision, creating a more robust system that maintains simplicity while improving security against impersonation attacks.

Inventive Principle:

Principle #5Merging (Combining)

2Speed

If only acoustic features are analyzed, then the processing speed is fast, but the authentication accuracy is insufficient to detect impersonation

Engineering Contradiction:

Improveprocessing speedVSAvoidauthentication accuracy

Core Design Contradiction:

SpeedVSMeasurement precision

Solution Approach 1:

The authentication system is segmented into two independent analysis pathways: acoustic feature analysis and linguistic feature analysis. Each pathway processes voice signals independently and generates its own score. This segmentation allows each model to specialize in detecting specific aspects of voice authenticity while maintaining overall processing efficiency, and the combined results improve detection accuracy for impersonation attempts.

Inventive Principle:

Principle #1Segmentation

3Measurement precision

If a dual-machine learning approach combining acoustic and linguistic models is used, then authentication accuracy is enhanced, but the system complexity increases

Engineering Contradiction:

Improveauthentication accuracyVSAvoidsystem complexity

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The patent introduces score combination mechanisms as intermediaries between the acoustic model and linguistic model outputs and the final authentication decision. The acoustic score and linguistic score are combined through weighted averaging or other fusion techniques, providing a structured way to integrate multiple analysis results. This intermediary layer manages the complexity of the dual-model system while maintaining high authentication accuracy through systematic score aggregation.

Inventive Principle:

Principle #24Intermediary (Mediator)

Data Source

PatentUS12573406B2Voice authentication based on acoustic and linguistic machine learning models

Publication Date: 2026.03.10 CISCO TECHNOLOGY INC

AI summary

In one example embodiment, acoustic characteristics of a user voice are analyzed by a first machine learning model of a processor. Linguistic patterns in the user voice are analyzed by a second machine learning model of the processor. The user is authenticated with respect to an authorized user by the processor based on analysis of the acoustic characteristics and the linguistic patterns of the user voice by the first and second machine learning models.