TL;DR
This paper introduces a novel black-box attack framework for Automatic Speech Recognition (ASR) systems using multi-objective evolutionary optimization. By simultaneously optimizing for acoustic similarity and text dissimilarity, the method generates highly effective adversarial audio samples that can significantly degrade ASR performance while remaining perceptually indistinguishable from the original audio.
What Problem Does This Paper Address?
The increasing ubiquity of Automatic Speech Recognition (ASR) systems in smart devices and voice assistants raises critical security concerns. These systems, while powerful, are vulnerable to adversarial attacks – subtle modifications to input audio that cause misinterpretation. Existing adversarial attack methods often fall into two categories: white-box attacks (requiring full model access, which is rare for commercial systems) or black-box attacks that may produce perceptible audio distortions or rely on specific model assumptions. This paper addresses the need for effective, practical black-box attacks on ASR systems that can fool them without introducing noticeable audio artifacts, thereby posing a significant threat to the security of voice-controlled systems.
Methods
The core of this work is a framework that frames adversarial audio generation as a multi-objective optimization problem. This approach aims to find a balance between two conflicting objectives:
- Maximizing Text Dissimilarity: Making the ASR system produce an incorrect transcription (either randomly for un-targeted attacks or a specific target text for targeted attacks).
- Minimizing Acoustic Dissimilarity: Ensuring the adversarial audio sample is perceptually very similar to the original audio.
The framework is algorithm-agnostic and can leverage different multi-objective evolutionary algorithms. The authors specifically implement and evaluate it using:
- Multi-Objective Genetic Algorithm (MOGA): A standard evolutionary algorithm for multi-objective optimization.
- Elitist Non-dominated Sorting Genetic Algorithm II (NSGA-II): A more advanced and widely used algorithm for multi-objective optimization.
The fitness of each candidate adversarial audio sample is evaluated based on two metrics:
- Acoustic Similarity: Measured by the Euclidean distance between Mel-Frequency Cepstral Coefficients (MFCCs) of the original and generated audio.
- Text Dissimilarity: Measured by the edit distance between the ASR's transcription of the original audio and the ASR's transcription of the generated adversarial audio (or a desired target text for targeted attacks).
The evolutionary process involves initialization, fitness evaluation, mating pool selection (using various schemes like rank-based and roulette wheel selection), crossover, and mutation, iteratively refining the population of adversarial samples until convergence.
Datasets
- Mozilla Common Voice dataset (CVS): Used for generating and evaluating adversarial samples. The authors selected 100 audio samples from this dataset for their experiments.
Key Insights and Ideas
- Multi-Objective Optimization for Adversarial Audio: Framing adversarial audio generation as a multi-objective problem is a novel approach. It allows for a more nuanced optimization that balances the need to fool the ASR with the crucial requirement of maintaining audio fidelity.
- Black-Box Effectiveness: The framework demonstrates strong performance in a black-box setting, meaning it doesn't require knowledge of the ASR model's internal architecture or weights, making it highly practical for attacking real-world systems.
- High Acoustic Similarity: A key finding is that the generated adversarial samples maintain a very high acoustic similarity (correlation coefficient of 0.97-0.98) with the original audio, making them difficult for humans to distinguish.
- Significant ASR Degradation: The attacks lead to substantial increases in Word Error Rate (WER), indicating the potency of the generated adversarial samples in disrupting ASR performance.
- Algorithm Agnosticism: The framework's ability to work with different evolutionary algorithms (MOGA and NSGA-II) highlights its flexibility and robustness.
Key Results
- Word Error Rate (WER) Increase: The adversarial attacks significantly increased the WER for both Deepspeech and Kaldi-ASR systems. For Deepspeech, WER increased by up to 980% in un-targeted attacks.
- Acoustic Similarity: The adversarial samples maintained a high acoustic similarity, with correlation coefficients (CC) of 0.98 and 0.97 for Deepspeech and Kaldi-ASR, respectively.
- Targeted Attack Success: In targeted attacks, the generated text was, on average, only 1.0-1.83 words away from the desired target text, demonstrating effective manipulation.
- Qualitative Evaluation: Human listening tests with 27 participants showed that:
- Humans transcribed the adversarial samples with a mean WER of 1.5, suggesting they couldn't easily detect differences.
- Over 80% of participants required little to no effort to comprehend the adversarial audio, indicating minimal audible noise.
- Participants could only distinguish between original and adversarial samples about 50% of the time, akin to random guessing.
Limitations and Future Work
- Transferability: While the paper briefly touches on transferability (testing samples generated for one ASR on another), it notes that while some common weaknesses exist, model-specific weaknesses also play a role. Further research is needed to understand and exploit this.
- Human Perception vs. ASR Performance: While the generated samples are perceptually similar to humans, the paper focuses on fooling the ASR. Future work could explore optimizing for human perception more directly.
- Computational Cost: Evolutionary algorithms can be computationally intensive. While the paper demonstrates effectiveness, optimizing the efficiency of the generation process could be a future direction.
- Exploration of Other ASR Systems: The current evaluation is on two specific ASR systems. Testing the framework on a wider variety of ASR architectures and commercial systems would be valuable.
- Defense Mechanisms: The paper implicitly highlights the need for robust defense mechanisms against such sophisticated adversarial attacks.
Summary updated