SyntheticPop: Attacking Speaker Verification Systems With Synthetic VoicePops

arXiv:2502.09553 · cs.CR, cs.LG · Submitted 2025-02-13 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Security Radio. Generated commentary on the latest security and cryptography papers.

Nadia: We're looking at the paper: "SyntheticPop: Attacking Speaker Verification Systems With Synthetic VoicePops".

Elias: Here's the gist of the paper. Voice Authentication (VA) systems are vulnerable to deepfake audio threats, and this paper proposes a novel attack method called SyntheticPop to target VA+VoicePop systems by embedding synthetic ‘pop’ noises into spoofed audio samples The SyntheticPop attack involves embedding synthetic ‘pop’ noises into spoofed audio samples, significantly degrading the model’s performance Our experiments demonstrate that VA+VoicePop achieves 69% accuracy under normal conditions, 37% accuracy when subjected to a baseline label flipping attack, and just 14% accuracy under our proposed SyntheticPop attack.

Nadia: System Vulnerability and Defense Mechanism Voice Authentication (VA) is a critical component of biometric security, primarily deployed in automated telecommunication systems like telephone banking and healthcare transcription.

Elias: VA systems face two primary attack vectors: physical and logical attacks.

Nadia: The VoicePop defense mechanism for VA can be viewed as a form of liveliness detection implemented in VA+VoicePop.

Title and authors: Nadia: So we're looking at this paper titled "SyntheticPop: Attacking Speaker Verification Systems With Synthetic VoicePops". It’s about how you can target those voice authentication systems that use things like VoicePops for liveliness detection.

Elias: Yeah, the authors are proposing a new attack method called SyntheticPop to mess with those specific systems. It seems they want to show how embedding synthetic pop noises into spoofed audio samples can seriously degrade the model's performance on VA+VoicePop setups.

Priya: What does that actually mean for the system's security? Is this just a theoretical problem, or is it something practical we need to worry about right now?

Nadia: It’s definitely something practical because they show how easy it is to make the model fail when you introduce these synthetic pops. They’ve done experiments showing that under normal conditions, VA+VoicePop gets sixty-nine percent accuracy, but with their SyntheticPop attack, that accuracy drops all the way down to just fourteen percent.

Elias: That drop from sixty-nine to fourteen is pretty significant for a security system. It shows that this specific type of noise injection can completely flip the classification outcome against the model’s intent when it’s dealing with those pop noises occurring in the one hundred hertz range, as mentioned on page two of their work.

Priya: But what about how robust is this defense mechanism itself? Does VA+VoicePop offer any protection against these kinds of adversarial data poisoning attempts they are describing?

Nadia: Well, the paper tests it against a couple of things. They show that VA+VoicePop is actually pretty resilient to a simple label flipping attack, which means if you just try to flip the labels on some samples in the training data, it doesn't hurt accuracy much; their attack success rate there was around sixty point nine six percent.

Elias: But they clearly show that resilience isn't absolute. They tested against their own SyntheticPop attack and found that the model’s performance suffered a lot, specifically a thirty-five point three two percent decrease in the true positive rate and an increase in the false negative rate by that same amount, according to page three of the paper.

Priya: So they’re saying this attack is particularly effective because it causes the model to misclassify many fakes as real, which sounds like a major privacy concern if those fakes are used for identity verification.

Nadia: Exactly. The core finding is that SyntheticPop attacks can cause the model to classify many spoofed samples as real, and the success rate for those attack samples was ninety-five percent, which is very high in their testing setup.

Elias: It seems like the noise they are injecting isn't just random static; it's engineered to specifically target how the system extracts features like delta one and delta two during those pop noise transitions, as described on page three.

Title and authors: Priya: I wonder if focusing only on these specific acoustic events means the system might still be vulnerable to other types of deepfake audio that don't rely on synthetic pops at all.

Nadia: That’s a valid point for future work, and they actually suggest some improvements in their paper to tackle that issue. They propose integrating a defense mechanism specifically targeting these synthetic pop noises to make the system more robust against them.

Elias: And they also suggest implementing a detection layer that analyzes those features like delta one and delta two during pop noise transitions to better distinguish genuine human speech from replay attacks, which should help capture "the naturalness of a human speaking into a microphone" as they put it on page two.

Priya: That sounds like a good practical step for improving the actual deployment, moving beyond just the theoretical attack description and focusing on how to filter out those false positives in real-time.

Nadia: Right. And from their perspective, they also think enhancing the training methodology by using synthetic minority over-sampling technique, or SMOTE, on the real audio class within VA+VoicePop systems could help when testing with more balanced datasets because they noted the dataset is heavily skewed towards spoofs on page three.c.

Elias: That would definitely help if you’re working with real-world data that isn't perfectly balanced, though I wonder what the computational cost would be to apply SMOTE in a system that has these complex feature extraction steps involving GFCC and STFT transforms.

Priya: The cost is worth it if the resulting model is genuinely more reliable when dealing with mixed datasets, especially since they found VA+VoicePop showed resilience to those simpler label flipping attacks earlier on.

Nadia: So, to wrap up this section, the main point from "SyntheticPop: Attacking Speaker Verification Systems With Synthetic VoicePops" is that while VA+VoicePop holds up against simple label flipping, it’s very susceptible to these targeted synthetic noise injections which drastically cut its accuracy down to fourteen percent.

Elias: And they conclude that this type of attack can cause a drastic reduction in the system's overall accuracy and suggests exploring how such attacks transfer to other modalities and biometrics in future research.

Priya: For me, it’s important to remember that while these specific adversarial attacks are nasty, we still need broader testing with datasets focused on either replay or deepfake audio, but not both, to really understand the full scope of these vulnerabilities.

Nadia: That makes sense. So we’ve seen how SyntheticPop works against VA+VoicePop and what defenses they suggest for that specific noise type. Next up, let's see what else is happening in the research space with another paper on Temporary Authority and permanent effects for LLM agents.

The paper's summary: Nadia: So we've seen how this SyntheticPop attack can drop accuracy down to fourteen percent on VA+VoicePop systems, so now we're looking at what the paper actually says about the attack itself.

Elias: Right, it basically says they created a new method called SyntheticPop that embeds these synthetic pop noises into spoofed audio samples and shows how that seriously degrades the model's performance.

Priya: What does embedding a pop noise into audio mean in practice for security? Is this something someone can actually use against a live system easily?

Nadia: It means they are injecting a specific type of noise—a synthetic pop—that targets how the VoicePop defense mechanism works, which is based on detecting those sudden shifts in energy and timing.

Elias: The paper details how they generate these samples by adding a subtle sine wave to the audio signal to make it sound like pop noise, and that's what causes the model to misclassify fakes as real.

Priya: So, the core issue isn't just any kind of fake audio; it’s specifically audio that has been engineered with these synthetic pop noises so it bypasses the detection logic.

Nadia: Exactly, and they show that this attack is very successful, with a ninety-five percent success rate across all their attack samples.

Elias: And while the defense system, VA+VoicePop, did hold up pretty well against a simple label flipping attack—that was only sixty point nine six percent success for that—it completely failed against the SyntheticPop method.

Priya: That contrast is important; it shows resilience to one kind of manipulation but total vulnerability to another specific type of adversarial input.

Nadia: Right, and the authors stress that this synthetic noise directly causes the model to degrade in classification, which results in a drastic reduction in overall VA+VoicePop accuracy.

Elias: They're hinting that this attack could potentially be used as a deterrent to the entire system because it introduces so much noisy data that other types of non-replay attacks might also get misclassified as accurate.

Priya: That’s a big implication for deployment; if your system starts getting flooded with this kind of synthetic noise, it might start trusting bad inputs too.

Nadia: It definitely points toward the need for better defenses that specifically target these types of engineered acoustic perturbations rather than just general spoofing detection.

Elias: They suggest that future work should look at how these SyntheticPop attacks transfer to other biometric modalities like fingerprint or iris verification systems, exploring the transferability of this noise perturbation.

Priya: It sounds like the next big hurdle is figuring out how to make a defense robust enough against this kind of specific, engineered noise.

The paper's improvements: Tom: So we’re looking at what the authors suggest as fixes for this SyntheticPop attack, and how those changes might actually make VA+VoicePop systems better at stopping these things.

Nadia: The paper proposes three main improvements to make the system more robust against these adversarial data poisoning attempts, starting with building a defense mechanism that specifically targets those synthetic pop noises.

Elias: That means they want to integrate a defense layer right into VA+VoicePop that’s designed to recognize and block audio samples where synthetic pop noises are embedded.

Priya: So instead of just trying to spot the general fake, they want a feature detector for this very specific noise artifact.

Nadia: Exactly, and they also suggest implementing a detection layer that analyzes those features like delta one and delta two during pop noise transitions to better distinguish genuine human speech from replay attacks.

Elias: That would be about capturing what makes human speech sound natural when it comes to those sudden acoustic events, which is crucial for filtering out background noises mistaken for pops.

Priya: From a privacy perspective, focusing on that naturalness helps ensure the system isn't fooled by just random noise patterns that happen to mimic a pop briefly.

Nadia: They also talk about enhancing the training methodology using something called synthetic minority over-sampling technique, or SMOTE, on the real audio class within VA+VoicePop systems.

Elias: That addresses the problem where the training data is heavily skewed towards spoofs, helping balance out that imbalance when testing with more balanced datasets.

Priya: Balancing the dataset helps show us what happens when we test against a more realistic mix of real and fake audio, which is what we need to trust the claimed result on.

Nadia: The authors also suggest developing an attack simulation framework that systematically tests how these SyntheticPop attacks transfer across different biometric modalities beyond just voice authentication.

Elias: So they want to see if this type of pop noise perturbation can be applied to other systems, like fingerprint or iris verification, to explore that transferability you mentioned earlier.

Priya: That’s a big leap; it moves the research from just fixing one problem in voice authentication to understanding how these specific acoustic vulnerabilities behave across different types of biometrics.

Nadia: The authors flag that their current testing with the ASVSpoof two thousand nineteen dataset has some limitations, so they suggest future studies should look at single attack types, like replay or deepfake, but not both.

Elias: So the paper points out that testing only one type of attack might give a clearer picture before trying to tackle everything at once.

Priya: It’s a practical constraint; you can't fix everything at once when you don't even know what the biggest threat is in your specific deployment environment.

Conclusion: Tom: So we've covered how the SyntheticPop attack works and what improvements the authors suggest to stop it, now we're wrapping up this piece on "SyntheticPop: Attacking Speaker Verification Systems With Synthetic VoicePops."

Nadia: Basically, the paper shows that this specific noise injection can completely break a VA+VoicePop system by making it classify many fakes as real and dropping its accuracy significantly.

Elias: It’s a clear demonstration of how targeted adversarial input, rather than just random noise, can exploit the way liveliness detection is implemented in those voice authentication systems.

Priya: What this means for the real world is that we need to be much more careful about what data we feed into these security models because they aren't immune to these specific kinds of engineered acoustic attacks.

Nadia: Right, and while they point out that the ASVSpoof two thousand nineteen dataset has some limitations, like it only focuses on one type of attack, the authors suggest future work needs to look at testing against replay audio or deepfake audio separately.

Elias: That makes sense; you need to isolate the vulnerability before you try to build a defense that covers every possible attack vector simultaneously.

Priya: It’s about focusing on one thing at a time, figuring out if a system can handle replay noise versus deepfake noise, rather than trying to solve both at once.

Nadia: So, the implication is that for anyone building these voice authentication systems, they need to specifically test against these engineered pop noises because simple defenses against general spoofing might not be enough.

Elias: I agree; it’s about understanding exactly what parameters break the proof so you can harden the system where it matters most.

Priya: It really comes down to ensuring that when we deploy these systems, we aren't just relying on a single defense mechanism that gets bypassed by something as simple as an embedded synthetic pop.

Eshaq Jamdar, Amith Kamath Belman

San Jose State University

cs.CR, cs.LG

Submitted: 2025-02-13

Updated: 2025-02-13

DOI: 10.1109/SVCC65277.2025.11133635

Code: https://github.com/MohamadMerchant/Voice-Authentication-and-Face-Recognition

License: http://creativecommons.org/licenses/by-sa/4.0/

Importance score: 73/100

The gist: Voice Authentication (VA) systems are vulnerable to deepfake audio threats, and this paper proposes a novel attack method called SyntheticPop to target VA+VoicePop systems by embedding synthetic…

Key concepts

Voice Authentication (VA)
VA is a biometric security feature used in systems like telephone banking or healthcare transcription. It verifies the identity of a speaker by analyzing their voice, making it critical for automated security processes.
Label Flipping Attack
This attack involves manipulating labeled data within the training set to confuse the system. The goal is to make the classifier incorrectly assign labels, such as misclassifying a genuine sample as a spoof or vice versa.
SyntheticPop Attack
This specific attack generates poisoned spoof samples by subtly adding a sine wave resembling 'pop' noises into audio. This noise perturbation is designed to disrupt the SVM classifier used in VA+VoicePop systems, causing significant performance degradation.

Terminology

Summary

Voice Authentication (VA) systems are vulnerable to deepfake audio threats, and this paper proposes a novel attack method called SyntheticPop to target VA+VoicePop systems by embedding synthetic ‘pop’ noises into spoofed audio samples The SyntheticPop attack involves embedding synthetic ‘pop’ noises into spoofed audio samples, significantly degrading the model’s performance Our experiments demonstrate that VA+VoicePop achieves 69% accuracy under normal conditions, 37% accuracy when subjected to a baseline label flipping attack, and just 14% accuracy under our proposed SyntheticPop attack. p. 0

Where the paper says it

We propose a novel attack method, which we refer to as SyntheticPop, designed to target the phoneme recognition capabilities of the VA+VoicePop system. The SyntheticPop attack involves embedding synthetic ‘pop’ noises into spoofed audio samples, significantly degrading the model’s performance.

System Vulnerability and Defense Mechanism

Voice Authentication (VA) is a critical component of biometric security, primarily deployed in automated telecommunication systems like telephone banking and healthcare transcription. VA systems face two primary attack vectors: physical and logical attacks. p. 0

Where the paper says it

It is primarily deployed in automated telecommunication systems, such as telephone banking, healthcare transcription, and smart home devices, where it acts as a secondary layer of authentication. VA systems face two primary attack vectors: physical and logical attacks [1].

The VoicePop defense mechanism for VA can be viewed as a form of liveliness detection implemented in VA+VoicePop. p. 0
Where the paper says it

Such detections allow greater confidence in authenticating users properly. The VoicePop defense mechanism for VA can be viewed as a form of liveliness detection implemented in VA+VoicePop [2].

The basis for VA+VoicePop can be traced back to the usage of phonemes as a way to perform liveliness detection on VAs. p. 0
Where the paper says it

The VoicePop defense mechanism for VA can be viewed as a form of liveliness detection implemented in VA+VoicePop [2]. The basis for VA+VoicePop can be traced back the usage of phonemes as a way to perform liveliness detection [4] on VAs.

Replicating the Defense Architecture

The authors recreated the VA+VoicePop system and tested it with large external datasets to address limitations in prior work. The methodology involves several steps to replicate the original system's logic:

  1. Extract pop noise timings and send them to GFCC filters to extract essential features. p. 1

    Where the paper says it

    The logic in Algorithm 1 explains the process of Pop Noise detection along with GFCC extraction of features to enable liveliness detection. For every audio file, pop noise timings are extracted, and these timings are then sent to GFCC filters to extract the essential features: • gfcc mean: average of pop noise magnitudes within the audio

  2. Calculate features such as gfcc mean, timing changes from pop noise to background (δ1), and the sudden shift in energy (δ2). p. 1

    Where the paper says it

    A mean of the overall magnitudes is calculated to represent one of the feature vectors to the Support Vector Machines (SVM). In addition, two other features are calculated: the timing changes from pop noise to background (δ1) and the sudden shift in energy in the recording (δ2).

  3. Use a Support Vector Machine (SVM) classifier to identify if the audio sample is from a human or a replay attack. p. 1

    Where the paper says it

    A replay attack, on the other hand, will have much lower values for both deltas as the noise captured by the microphone cannot pick up on these details due to the audio sample being played from a speaker, which causes these details to be silenced or less pronounced. Once the GFCC features are extracted, a SVM classifier is used to identify if the audio sample is from a human or a replay attack.

Attack Design and Generation

The research investigates two types of attacks:

  1. Label Flipping Attack: This straightforward method targets labeled data within a population to negate the labels. p. 3

    Where the paper says it

    A high level overview of the attack strategies is presented in 4. 1) Label Flipping Attack: The label flipping attach is a straightforward method that can target a certain amount of labeled data within a population and negate the labels [7].

  2. SyntheticPop Attack: SVMs can be targeted by altering training data in hopes that these modifications are enough to disrupt the classification process. p. 3

    Where the paper says it

    These types of attacks are commonly used with Binary Classification problems, making it a good candidate for VA+VoicePop. 2) SyntheticPop Attack: SVMs can be targeted by altering training data in the hopes that these modifications are enough to disrupt the classification process [8].

    The SyntheticPop attack generates poisoned spoof samples by adding a subtle sine wave to the audio signal to resemble pop noise. p. 4
    Where the paper says it

    of energy that is below or equal to 100 Hz on lower bands, a subtle sine wave can be added to resemble pop noise in a fake audio sample. The control variables in the algorithm dictate the amplitude and frequency values of the signal.

Experimental Results and Findings

The results show that VA+VoicePop showed resilience to simple label flipping attacks but was susceptible to SyntheticPop. Specifically:

** Label Flipping Attack:**

  • The attack success rate (ASR) was 60.96%.

  • Poisoning 20% of the samples yielded an overall accuracy of 37.21%, a -0.31 difference. p. 4

    Where the paper says it

    Table II’s 20% poison with label flipping are similar to Table I’s even train result. Calculating accuracy difference based on Table I’s even train row, poisoning 20% of the samples yields an overall accuracy of 37.

  • The Attack Success Rate (ASR) shows a 60.96% success rate. p. 0

    Where the paper says it

    The SyntheticPop attack involves embedding synthetic ‘pop’ noises into spoofed audio samples, significantly degrading the model’s performance. We achieve an attack success rate of over 95% while poisoning 20% of the training dataset.

** SyntheticPop Attack:**

  • The model’s accuracy has fallen to an accuracy 14.01%, with a TPR decrease of over 35. p. 5

    Where the paper says it

    Considering the accuracy in Table II in comparison to the label flip attack statistics and even training from Table I (as both results are very similar to one another), the model’s accuracy has fallen to an accuracy 14. 01%, with a TPR decrease of over 35.

    32% and the FNR increased by the same amount. p. 5
    Where the paper says it

    01%, with a TPR decrease of over 35. 32% and the FNR increased by the same amount.

  • SyntheticPop attacks caused the model to classify many fakes as real.

  • The ASR is 95% of all attack samples made it through, indicating a very high success rate.

Conclusion and Future Work

The key insights are that VA+VoicePop showed resilience to simple label flipping attacks but was susceptible to SyntheticPop. SyntheticPop attacks deployed directly cause the model to degrade in classification and ultimately result in a drastic reduction in VA+VoicePop accuracy. p. 5

Where the paper says it

Under label flipping attacks, fake samples with flipped labels do not cause a large difference in VA+VoicePop accuracy. b) SyptheticPop attacks deployed directly, cause the model to degrade in classification and ultimately resulted in a drastic reduction in VA+VoicePop accuracy.

This type of attack proved to be a deterrent to the overall system and could have possibly allowed other types of non-replay attacks to be classified as accurate due to this noisy data. p. 5
Where the paper says it

b) SyptheticPop attacks deployed directly, cause the model to degrade in classification and ultimately resulted in a drastic reduction in VA+VoicePop accuracy. This type of attack proved to be a determent to the overall system and could have possibly allowed other types of non-replay attacks to be classified as accurate due to this noisy data.

Further studies with datasets focused on single type of attacks, like either replay or deepfake, but not both, would help found a clearer understanding of these vulnerabilities in VA systems. p. 5
Where the paper says it

The use of ASVSpoof 2019 dataset presents subtle limitations to our work. p. 5

Where the paper says it

This is an important finding, as many Voice Recognition systems that can differentiate between human, bots, or replay attacks can be vulnerable to these issues depending on where they receive such data. The use of ASVSpoof 2019 dataset presents subtle limitations to our work.

Further studies with datasets focused on single type of attacks, like either replay or deepfake, but not both, would help found a clearer understanding of these vulnerabilities in VA systems.

The use of ASVSpoof 2019 dataset presents subtle limitations to our work. Further testing incorporating diffusion based deepfake audio with pop noises would help analyze the effect of cloning samples on a VA+VoicePop system. p. 5
Where the paper says it

Further studies with datasets focused on single type of attacks, like either replay or deepfake, but not both, would help found a clearer understanding of these vulnerabilities in VA systems. Attack testing incorporating diffusion based deepfake audio with pop noises would help analyze the effect of cloning samples on a VA+VoicePop system.

The authors suggest that implementing approaches like SyntheticPop described in this study to explore the transferability of such attacks to different modalities and biometrics. p. 5
Where the paper says it

Although some studies show generative models can sometimes make audio sound too unnatural and too clean [13]. It would be intriguing to implement approaches like SyntheticPop described in this study to explore the transferability of such attacks to different modalities and biometrics.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improve VA+VoicePop robustness against adversarial data poisoning by integrating a defense mechanism specifically targeting synthetic pop noises, as demonstrated by SyntheticPop. This enhanced system can maintain higher accuracy when subjected to attacks that embed synthetic 'pop' noises into spoofed audio samples, potentially mitigating the degradation observed under the original attack.

  2. Implement a detection layer that analyzes features like delta1 and delta2 during pop noise transitions to distinguish between genuine human speech and replay attacks more effectively. This will allow the system to better capture the naturalness of a human speaking into a microphone, as described in Section II, which is crucial for filtering out background noises mistaken for pops.

  3. Enhance the training methodology by utilizing synthetic minority over-sampling technique (SMOTE) on the real audio class within VA+VoicePop systems, combined with balanced weight configurations. This will address the issue where the dataset is heavily skewed towards spoofs and improve model performance when testing with more balanced datasets, as suggested in Section III.C.

  4. Develop an attack simulation framework that systematically tests the transferability of SyntheticPop attacks across different biometric modalities beyond voice authentication. This allows researchers to explore the transferability of such attacks to different modalities and biometrics, as mentioned in the conclusion, by applying pop noise perturbations to other speaker recognition systems like fingerprint or iris verification.

Abstract

Voice Authentication (VA), also known as Automatic Speaker Verification (ASV), is a widely adopted authentication method, particularly in automated systems like banking services, where it serves as a secondary layer of user authentication. Despite its popularity, VA systems are vulnerable to various attacks, including replay, impersonation, and the emerging threat of deepfake audio that mimics the voice of legitimate users. To mitigate these risks, several defense mechanisms have been proposed. One such solution, Voice Pops, aims to distinguish an individual's unique phoneme pronunciations during the enrollment process. While promising, the effectiveness of VA+VoicePop against a broader range of attacks, particularly logical or adversarial attacks, remains insufficiently explored. We propose a novel attack method, which we refer to as SyntheticPop, designed to target the phoneme recognition capabilities of the VA+VoicePop system. The SyntheticPop attack involves embedding synthetic "pop" noises into spoofed audio samples, significantly degrading the model's performance. We achieve an attack success rate of over 95% while poisoning 20% of the training dataset. Our experiments demonstrate that VA+VoicePop achieves 69% accuracy under normal conditions, 37% accuracy when subjected to a baseline label flipping attack, and just 14% accuracy under our proposed SyntheticPop attack, emphasizing the effectiveness of our method.

Sources

Related papers