TL;DR
This paper introduces a novel "timbre-reserved adversarial attack" for speaker identification (SID) systems. By integrating adversarial constraints into voice conversion (VC) models, they generate fake audio that not only fools SID systems but also retains the target speaker's natural timbre and can be customized for specific text, making attacks more sophisticated and harder to detect.
What Problem Does This Paper Address?
Speaker identification (SID) systems, crucial for biometric authentication, are vulnerable to various attacks. Existing attacks fall into two main categories: spoofing and adversarial. Spoofing attacks, like voice conversion or synthesis, aim to mimic a target speaker's voice but often fail to exploit the underlying vulnerabilities of SID models, leading to suboptimal attack performance. Adversarial attacks, on the other hand, cleverly perturb audio to fool SID models but often result in unnatural-sounding audio that sacrifices the target speaker's timbre and can be easily detected by humans. This leaves a gap for attacks that are both effective against SID systems and perceptually convincing, preserving the target speaker's identity and naturalness.
Methods
The core innovation is a "timbre-reserved adversarial attack" achieved by modifying the training process of voice conversion (VC) models. Instead of directly adding adversarial noise to the final audio, the authors inject adversarial constraints during the VC model's training. This is done by using a speaker classifier to evaluate the generated audio's predicted speaker identity. If the classifier doesn't identify the target speaker, an adversarial perturbation is applied to the VC model's internal representations (Mel-spectrogram, latent representation, or reconstructed waveform) to guide it towards producing audio that would be classified as the target speaker. This process is optimized to minimize the error of the SID system while also preserving the original reconstruction loss of the VC model. This approach ensures that the generated audio is not only adversarial but also retains the desired timbre and can be controlled for specific text content.
Datasets
- AISHELL-3: Used for training the speaker identification (SID) model and voice conversion (VC) models.
- AISHELL-1: Used for evaluating the performance of the SID model.
- Audio Deepfake Detection (ADD) challenge dataset: Used to evaluate the proposed timbre-reserved adversarial attack. This dataset is noted as being specifically designed for deepfake audio attack and detection.
Key Insights and Ideas
- Adversarial training of VC models is more effective than post-hoc perturbation: Injecting adversarial constraints during VC training allows the model to learn to generate adversarial audio inherently, leading to better quality and more natural-sounding results compared to simply adding noise to already generated audio.
- Targeting SID model vulnerabilities while preserving timbre is achievable: The proposed method successfully balances the need to exploit SID model weaknesses with the human perception requirement of maintaining the target speaker's timbre.
- Flexibility in adversarial constraint placement: The authors demonstrate that adversarial constraints can be effectively applied at different representation levels within the VC model (Mel-spectrogram, latent representation, or waveform), offering flexibility in implementation.
- Timbre and text customization are key for realistic attacks: The ability to control both the speaker's timbre and the spoken text makes the generated fake audio more convincing and useful for specific attack scenarios.
Key Results
- Improved Attack Success Rate: The proposed timbre-reserved adversarial attack strategies achieved significantly higher attack success rates compared to a vanilla VC model. For instance, the "Waveform" strategy achieved 66.30% attack success, a substantial improvement over the vanilla VC's 29.60%.
- Comparable to Direct Perturbation: The attack success rates of the proposed methods were comparable to, though slightly lower than, directly adding adversarial perturbation to VC-generated audio (VC+adv), which achieved 76.50%. This is attributed to the adversarial constraint being applied during training rather than directly to the final output.
- High Deception Success Rate (DSR): The generated fake audios demonstrated a high ability to fool detection systems, with DSRs often exceeding 90% for the proposed methods, comparable to ADD challenge participating teams.
- Superior Audio Quality: Objective (o-MOS) and subjective (CMOS) evaluations showed that the proposed methods produced fake audio with better quality and similarity than the VC+adv method, and comparable to the vanilla VC model, indicating that adversarial constraints did not degrade audio quality.
- Low Character Error Rate (CER): The generated fake audios maintained low CERs, confirming that the text content was preserved and intelligible.
Limitations and Future Work
- SID Model Vulnerability: The paper notes that the adversarial constraint on latent representation was harder to optimize, resulting in a lower attack success rate compared to other strategies. This suggests that the effectiveness of the adversarial constraint might depend on the specific representation chosen and its proximity to the SID model's input.
- Future Work: The authors suggest that their work can be extended to other types of adversarial attacks and that further research into optimizing adversarial constraints at different representation levels could yield even better results. The ability to customize text and timbre also opens doors for more targeted and sophisticated attacks in various scenarios.
Summary updated