Boosting the Transferability of Audio Adversarial Examples with Acoustic Representation Optimization

TL;DR

This work tackles the challenge of generating transferable audio adversarial examples for unseen ASR systems by proposing Acoustic Representation Optimization (ARO). ARO guides perturbations toward low-level acoustic representations extracted from SRMs, rather than model-specific features, via a cosine-similarity loss that complements traditional adversarial objectives. The method is plug-and-play, enabling integration with existing attacks, and experiments across multiple ASR models and SRMs show substantial gains in transferability with minimal audio-quality degradation. The findings indicate that lower-layer acoustic features generalize better across architectures, offering a practical enhancement for black-box adversarial scenarios.

Abstract

With the widespread application of automatic speech recognition (ASR) systems, their vulnerability to adversarial attacks has been extensively studied. However, most existing adversarial examples are generated on specific individual models, resulting in a lack of transferability. In real-world scenarios, attackers often cannot access detailed information about the target model, making query-based attacks unfeasible. To address this challenge, we propose a technique called Acoustic Representation Optimization that aligns adversarial perturbations with low-level acoustic characteristics derived from speech representation models. Rather than relying on model-specific, higher-layer abstractions, our approach leverages fundamental acoustic representations that remain consistent across diverse ASR architectures. By enforcing an acoustic representation loss to guide perturbations toward these robust, lower-level representations, we enhance the cross-model transferability of adversarial examples without degrading audio quality. Our method is plug-and-play and can be integrated with any existing attack methods. We evaluate our approach on three modern ASR models, and the experimental results demonstrate that our method significantly improves the transferability of adversarial examples generated by previous methods while preserving the audio quality.

Figures (4)

• Figure 1: Illustration of the transferability of the attack method with/without acoustic representation optimization.
• Figure 2: Overview of acoustic feature optimization integrated with traditional attack methods.
• Figure 3: The variation of SRoA with the depth of representation layers in different models.
• Figure 4: Impact of hyper-parameter $\beta$.