TL;DR
This paper explores a more realistic attack scenario for text-dependent speaker verification (TD-SV) systems, where attackers don't need the genuine speaker's password but can use their normal speech to craft adversarial noise. They demonstrate that even state-of-the-art TD-SV systems are vulnerable to these attacks, achieving a significant success rate.
What Problem Does This Paper Address?
Speaker Verification (SV) systems authenticate individuals based on their voice. These systems are broadly categorized into text-independent (TI-SV) and text-dependent (TD-SV). TI-SV systems verify only the speaker's voice characteristics, while TD-SV systems additionally verify that the speaker is saying a specific password. While adversarial attacks on TI-SV systems have been studied, a key assumption in many of these studies is that the attacker has access to the genuine speaker's enrollment speech. This is considered unrealistic, as having this access would allow direct bypass of the system.
This paper argues that TD-SV systems present a more practical attack surface. In a real-world scenario, an attacker might not have the genuine speaker's password but could likely obtain recordings of their normal speech. The problem this paper addresses is the feasibility of generating adversarial noise that, when added to the genuine speaker's normal speech, can fool a TD-SV system into accepting an imposter or rejecting a genuine user, specifically by bypassing the password requirement.
Methods
The core of the paper's methodology lies in generating adversarial noise specifically for TD-SV systems. Unlike previous TI-SV attacks, this approach does not require the genuine speaker's enrollment speech (i.e., the correct password). Instead, it leverages natural conversations from the genuine speaker, referred to as "Target Wrong" (TW) utterances.
The proposed attack involves:
- Target Wrong Utterances: Collecting multiple utterances from the genuine speaker that do not contain the correct password.
- Attacker Utterance: Using an utterance spoken by an imposter. This utterance can either use the correct password (Imposter Correct) or an incorrect password (Imposter Wrong).
- Adversarial Noise Generation: The key novelty is minimizing an objective function that aims to make the imposter's utterance, when combined with adversarial noise, appear similar to the genuine speaker's TW utterances. This is achieved using a gradient descent approach, iteratively updating the adversarial noise
v[n]to satisfyfo(stw[n], s[n] + v[n]) > τ + kfor all TW utterances, wherefois the TD-SV system's similarity measure,τis the system's decision threshold, andkis an attack confidence parameter. The noise is constrained to a small norm. - Attack Execution: The generated adversarial noise
v[n]is added to the imposter's utterances[n]to create an adversarial signala[n] = s[n] + v[n]. This adversarial signal is then presented to the TD-SV system.
The TD-SV system itself is described as a deep neural network classifier trained on speaker-password pairs. It uses a pre-trained WavLM model as a feature extractor, followed by a Multi-Head Factorized Attentive (MHFA) pooling layer and a segment-level encoder to produce speaker-password posteriors.
Datasets
- Training Data: The TD-SV system was trained using the Track 1 training data from the TDSV 2024 challenge. This dataset comprises multiple repetitions of 10 passwords (5 English, 5 Persian) spoken by 1620 speakers.
- Evaluation Data: The RedDots database was used for evaluating the performance and adversarial vulnerability of the TD-SV system. Specifically, the "Common Pass-phrases Text-Dependent" evaluation set was used, which includes 10 identical passwords spoken by 62 different speakers (49 male, 13 female).
Key Insights and Ideas
- Realistic Attack Scenario: The paper successfully shifts the focus of adversarial attacks from an unrealistic "white-box" scenario for TI-SV to a more plausible "grey-box" scenario for TD-SV, where attackers leverage readily available natural speech.
- Password Vulnerability Varies: Not all passwords are equally susceptible to adversarial attacks. The study found that "Okay Google" was more robust, while "Birthday parties have cupcakes and ice cream" was highly vulnerable, suggesting that linguistic properties of passwords might play a role.
- TD-SV Systems Remain Vulnerable: Despite advancements in TD-SV, the proposed adversarial attack demonstrates significant vulnerability, even when the adversarial noise is added at a relatively high Signal-to-Noise Ratio (SNR).
Key Results
- Overall Attack Success Rate: The proposed adversarial attack achieved an average attack success rate of 64.28% when targeting "Target Wrong" utterances (i.e., the attacker's speech contained the correct speaker but the wrong password).
- Performance with Different Attacker Utterances:
- Target Wrong (TW): 64.28% success rate (SNR 30.30 dB)
- Imposter Correct (IC): 60.52% success rate (SNR 35.01 dB)
- Imposter Wrong (IW): 42.12% success rate (SNR 32.37 dB)
- TD-SV System Performance: The WavLM-based TD-SV system showed lower Equal Error Rates (EERs) compared to older i-vector/GMM and i-vector/HMM systems, indicating its state-of-the-art performance.
Limitations and Future Work
- White-Box Assumption: The current attack is a gradient-based white-box attack, meaning it requires detailed knowledge of the TD-SV system's architecture and parameters. This limits its applicability in truly black-box scenarios.
- Password-Specific Vulnerability: While the paper identifies varying password vulnerabilities, the underlying reasons (e.g., phoneme composition) are not fully explored.
- Future Work: The authors suggest exploring adversarial noise generation using transferable black-box attacks or natural evolution strategies for black-box gradient estimation. Further investigation into the linguistic properties of passwords and their impact on adversarial robustness is also planned.
Summary updated