Snooow1029/advwave-adversarial-audio

AdvWave Adversarial Audio (in-house reproduction)

Adversarial audio generated with the AdvWave waveform-suffix attack (Kang et al.) for authorized AI-safety / adversarial-robustness research. Each clip is an AdvBench harmful prompt (TTS) with an optimized adversarial waveform suffix that induces a target audio-LM to comply.

⚠️ Research artifact. Intended solely for defensive research and red-teaming evaluation of Large Audio-Language Models. Do not use to elicit harmful content.

Subsets

• qwen2_audio/ — targeted at Qwen2-Audio-7B-Instruct (AdvBench-520, CSV affirmative target). AdvWave waveform-suffix attack, tight early-stop CE<0.1 (~360 epochs, 514/520 converged, CE mean 0.074). Reproduces the AdvWave paper within ~3pp — undefended Qwen2-Audio ASR-W 0.860 / HEx-PHI ASR-L 0.852 (paper 0.891 / 0.884).

• llama_omni/ — targeted at Llama-3.1-8B-Omni (AdvBench-520, CSV affirmative target). AdvWave waveform-suffix attack, tight early-stop CE<0.1 (mean CE 0.063, 516/520 converged). Model-specific optimization: Omni-native llama_3 conversation template + OpenAI-Whisper log-mel (OmniMel) front-end and Omni forward path — NOT the Qwen2 template. Undefended Llama-3.1-8B-Omni: HEx-PHI ASR-L 0.619 (keyword non-refusal 0.594).

Each subset: audio/<id>.wav + manifest.json (question, target_model, attack convergence stats, target used).