Abstract
Existing audio adversarial attack methods suffer from poor transferability, primarily due to insufficient exploration of model decision mechanisms and overreliance on heuristic-driven algorithm design. This paper aims to alleviate this gap. Specifically, through observations across three mainstream audio tasks (Automatic Speech Recognition, Speaker Verification, and Keyword Spotting), we reveal that these models primarily rely on local temporal features—inputs with time shuffled retain 83.7% of original accuracy. The SHAP-based visualization further validated that time shuffle leads to a significant shift in the salient regions of the model, but the samples can still be correctly identified, indicating the presence of redundant features that can affect decision-making. Inspired by these findings, we propose Time-Shuffle (TS) adversarial attack (including segments-based TS and phoneme-level-based TS-p). This method divides audio or phonemes into segments, randomly shuffles them, and computes gradients on the shuffled structure. By forcing perturbations to exploit transferable local temporal features and reduce overfitting to source-specific patterns, TS/TS-p inherently enhances transferability. As a model-agnostic framework, TS/TS-p can seamlessly integrate with existing attack methods. Comprehensive experiments demonstrate that TS-p achieved SOTA and boosts transferability by about 23%/14.7%/6.3% on ASR/ASV/KWS.
Introduction
Adversarial attacks—imperceptible perturbations designed to deceive AI models—pose a significant security threat to voice-activated systems. A primary challenge in this domain is transferability: an attack optimized for a specific "white-box" model often fails when deployed against an unknown "black-box" system. Previous attempts to improve transferability have relied on ensemble methods or generic data augmentations, which are computationally expensive and fail to address the underlying decision-making logic of audio Deep Neural Networks (DNNs). This research investigates the fundamental mechanisms of audio models to develop a more robust, transferable attack framework.
The Discovery of Temporal Insensitivity
The researchers began by investigating how audio models process temporal information. By dividing audio clips into segments and randomly shuffling their order, they observed that models for Automatic Speech Recognition (ASR), Speaker Verification (ASV), and Keyword Spotting (KWS) retained approximately 83.7% of their original accuracy. This suggests that these models are "temporally insensitive," relying more on local acoustic textures and sub-phonemic features than on global temporal structure.
Visualizing Feature Redundancy
To validate this, the authors used SHAP (SHapley Additive exPlanations) to visualize the salient regions—the parts of the input the model "pays attention" to.
As seen in the visualizations, while the original audio has specific regions of high importance, shuffling the time frames causes the model to pivot its attention to other redundant local features while still reaching the correct classification.
This redundancy is a critical vulnerability: if a model can find the "correct" answer in multiple parts of the audio, an attacker can exploit this by ensuring their perturbation is effective across all those redundant regions.
Methodology: The Time-Shuffle (TS) Attack
Inspired by the observation that models focus on local features, the authors proposed the Time-Shuffle (TS) attack framework. The process involves:
- Segmentation: The audio is divided into segments (TS) or specific phonemes (TS-p).
- Shuffling: During the attack generation phase, these segments are randomly rearranged.
- Gradient Aggregation: Gradients are calculated across multiple shuffled versions of the audio rather than just the original sequence.
- Optimization: This forces the adversarial perturbation to focus on local temporal features that are common across different models, effectively preventing the attack from "overfitting" to the specific global structure of the source model.
Experimental Results
The researchers evaluated the method against nine state-of-the-art models, including wav2vec, Conformer, and ECAPA-TDNN, using benchmarks like LibriSpeech and Google Speech Command V2.
The phoneme-level version (TS-p) achieved new state-of-the-art results in transferability:
- ASR: Boosted transferability (Word Error Rate) by 23%.
- ASV: Improved transferability (Cosine Similarity) by 14.7%.
- KWS: Increased the Transfer Attack Success Rate by 6.3%.
Conclusion
Time-Shuffle represents a significant advancement in audio adversarial research. By exploiting the inherent temporal insensitivity and feature redundancy of audio DNNs, TS-p creates perturbations that are far more generalized and effective against black-box systems. As a model-agnostic booster, it can be seamlessly integrated with existing attack methods to enhance their performance across a wide range of audio tasks.
Summary updated