When Good Sounds Go Adversarial: Jailbreaking Audio-Language Models with Benign Inputs
Fuente:
arXiv
Enregistré dans:
| Auteurs principaux: | , , , , , , , |
|---|---|
| Format: | Preprint |
| Publié: |
2025
|
| Sujets: | |
| Accès en ligne: | |
| Tags: |
Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
|
| _version_ | 1866914304665583616 |
|---|---|
| author | Dingeto, Hiskias Kwon, Taeyoun Choi, Dasol Kim, Bodam Lee, DongGeon Park, Haon Lee, JaeHoon Shin, Jongho |
| author_facet | Dingeto, Hiskias Kwon, Taeyoun Choi, Dasol Kim, Bodam Lee, DongGeon Park, Haon Lee, JaeHoon Shin, Jongho |
| contents | As large language models (LLMs) become increasingly integrated into daily life, audio has emerged as a key interface for human-AI interaction. However, this convenience also introduces new vulnerabilities, making audio a potential attack surface for adversaries. Our research introduces WhisperInject, a two-stage adversarial audio attack framework that manipulates state-of-the-art audio language models to generate harmful content. Our method embeds harmful payloads as subtle perturbations into audio inputs that remain intelligible to human listeners. The first stage uses a novel reward-based white-box optimization method, Reinforcement Learning with Projected Gradient Descent (RL-PGD), to jailbreak the target model and elicit harmful native responses. This native harmful response then serves as the target for Stage 2, Payload Injection, where we use gradient-based optimization to embed subtle perturbations into benign audio carriers, such as weather queries or greeting messages. Our method achieves average attack success rates of 60-78% across two benchmarks and five multimodal LLMs, validated by multiple evaluation frameworks. Our work demonstrates a new class of practical, audio-native threats, moving beyond theoretical exploits to reveal a feasible and covert method for manipulating multimodal AI systems. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2508_03365 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | When Good Sounds Go Adversarial: Jailbreaking Audio-Language Models with Benign Inputs Dingeto, Hiskias Kwon, Taeyoun Choi, Dasol Kim, Bodam Lee, DongGeon Park, Haon Lee, JaeHoon Shin, Jongho Sound Artificial Intelligence Cryptography and Security Audio and Speech Processing As large language models (LLMs) become increasingly integrated into daily life, audio has emerged as a key interface for human-AI interaction. However, this convenience also introduces new vulnerabilities, making audio a potential attack surface for adversaries. Our research introduces WhisperInject, a two-stage adversarial audio attack framework that manipulates state-of-the-art audio language models to generate harmful content. Our method embeds harmful payloads as subtle perturbations into audio inputs that remain intelligible to human listeners. The first stage uses a novel reward-based white-box optimization method, Reinforcement Learning with Projected Gradient Descent (RL-PGD), to jailbreak the target model and elicit harmful native responses. This native harmful response then serves as the target for Stage 2, Payload Injection, where we use gradient-based optimization to embed subtle perturbations into benign audio carriers, such as weather queries or greeting messages. Our method achieves average attack success rates of 60-78% across two benchmarks and five multimodal LLMs, validated by multiple evaluation frameworks. Our work demonstrates a new class of practical, audio-native threats, moving beyond theoretical exploits to reveal a feasible and covert method for manipulating multimodal AI systems. |
| title | When Good Sounds Go Adversarial: Jailbreaking Audio-Language Models with Benign Inputs |
| topic | Sound Artificial Intelligence Cryptography and Security Audio and Speech Processing |
| url | https://arxiv.org/abs/2508.03365 |