CodecAttack: Robust Adversarial Perturbations for Audio LLMs

Researchers have developed CodecAttack, a new adversarial attack method for Audio Large Language Models (Audio LLMs) that remains effective even after codec compression preprocessing. Unlike prior attacks that perturb the audio waveform—which can be detected and removed by codec compression—CodecAttack optimizes perturbations within a neural audio codec's continuous latent space. The attack exploits the codec's own compression channel, which discards waveform perturbations but transmits those crafted in its latent space. To enhance robustness across real-world compression channels, the method applies multi-bitrate straight-through Expectation-over-Transformation (EoT) without modifying the target model. The attack was tested across three realistic Audio LLM deployment scenarios and three target models, demonstrating consistent effectiveness. This work highlights a critical vulnerability in current defenses against adversarial attacks on audio AI systems.