SoundBreak: Studying vulnerabilities exposed by Audio Adversarial Attacks on Trimodal Models
Files
TR Number
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Foundation models that jointly process audio, visual, and textual inputs have demonstrated strong performance across a wide range of reasoning and generation tasks. Despite these advances, the robustness of such systems against adversarial manipulation remains insufficiently explored. This thesis presents SoundBreak, a comprehensive study of untargeted adversarial attacks that perturb only the audio modality in trimodal audio-video-language models. We introduce six base attack objectives and a combined objective that target different components of the multimodal pipeline, including audio encoder representations, cross-modal attention distributions, internal hidden-state dynamics, and output token likelihoods.
Extensive experiments conducted across four state-of-the-art multimodal models and multiple benchmark datasets show that audio-only perturbations are sufficient to induce substantial performance degradation, achieving attack success rates of up to 96%. The combined objective achieves the highest attack success rate, while the encoder-space objective provides the clearest high-effectiveness, low-distortion trade-off, reaching 89.12% ASR with LPIPS 0.08 and SI-SNR -1.77 dB. The lowest-distortion objectives achieve LPIPS 0.06 and SI-SNR around 0 dB, but with lower ASR.
Our analysis further reveals that prolonged optimization contributes more significantly to attack effectiveness than increased dataset scale, while transferability across models and audio encoders remains limited. Additionally, automatic speech recognition systems such as Whisper are found to respond primarily to overall distortion magnitude, reaching attack success rates exceeding 97% under high-distortion objectives. Collectively, these findings expose a previously underexplored single-modality vulnerability in multimodal systems and highlight the need for robust defense mechanisms based on cross-modal consistency and agreement verification.