Biometric Speaker Verification Playback Attack Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Text-dependent biometric speaker verification systems are vulnerable to playback attacks, where a recorded passphrase is played back to gain unauthorized access, and existing solutions either increase error rates or require more resource-intensive methods.
Innovation Solution
A method and system that prepare a spectrogram of recorded speech, normalize it, apply high-pass digital filtering, extract local maxima pairs, and calculate similarity to detect pre-recorded biometric information, using Hamming windowing and Gaussian masking to enhance security and efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If text-dependent biometric speaker verification systems are used, then verification accuracy is improved, but vulnerability to playback attacks increases
Solution Approach 1:
The system performs preliminary analysis of the speech signal during the verification process itself, extracting spectral peaks and computing the peakmap feature from the incoming speech signal before making the verification decision. This preliminary feature extraction enables playback detection without requiring separate pre-processing of reference recordings, resolving the contradiction by integrating security checks into the verification flow.
Solution Approach 2:
The invention extracts specific spectral features (spectral peaks) from the speech signal to create a compact peakmap representation. By taking out only the most relevant spectral characteristics rather than analyzing the entire signal, the system achieves effective playback detection while maintaining computational efficiency and verification accuracy.
2Object-affected harmful factors
If text-prompted systems are used to prevent playback attacks, then playback attack resistance is improved, but error rates increase
Solution Approach 1:
The peakmap-based detection method serves multiple functions: it detects playback attacks while simultaneously providing speaker verification capabilities. The same spectral peak extraction process used for playback detection also captures speaker-specific characteristics, allowing the system to maintain low error rates while achieving playback attack resistance.
Solution Approach 2:
The system changes the parameter space by transforming speech signals into a peakmap representation based on spectral peak locations and energies. This parameter transformation enables the system to detect playback attacks through spectral analysis while maintaining speaker verification accuracy, avoiding the high error rates associated with text-prompted systems.
3Measurement precision
If comprehensive spectral analysis is performed for playback detection, then detection accuracy is improved, but computational complexity increases
Solution Approach 1:
The system extracts only the most salient spectral features (the N highest spectral peaks) from the speech signal to create the peakmap. By taking out only these critical features rather than analyzing the complete spectral content, the system achieves high detection accuracy while significantly reducing computational complexity compared to comprehensive spectral analysis.
Solution Approach 2:
The invention applies partial action by analyzing only the N highest spectral peaks rather than the entire spectrum. This selective approach provides sufficient detection accuracy for playback attacks while keeping computational requirements manageable, avoiding the excessive complexity of full spectral analysis.
Data Source
Figure 1
Figure 2
Figure 3A~3B
AI summary
A method for detecting a recorded biometric information, the method comprising preparing (201) a spectrogram S n,k of the recorded speech biometric information, wherein the spectrogram S n,k is an N x K matrix, where S n,k is a log-amplitude of the spectrogram at a given time frame n and frequency bin k; the method being characterized in that it further comprises the steps of: normalizing (202) the spectrogram by its mean spectral value; subjecting (203) the spectrogram to high-pass digital filtering; extracting (204) a local maxima pair; and calculating (205) a similarity by comparing the output of the local maxima pair extraction with a previously stored data.