Y. Wang

Bocheng Chen, Guangjing Wang, Hanqing Guo et al.

Recent advances in natural language processing and machine learning have led to the development of chatbot models, such as ChatGPT, that can engage in conversational dialogue with human users. However, the ability of these models to generate toxic or harmful responses during a non-toxic multi-turn conversation remains an open research question. Existing research focuses on single-turn sentence testing, while we find that 82\% of the individual non-toxic sentences that elicit toxic behaviors in a conversation are considered safe by existing tools. In this paper, we design a new attack, \toxicbot, by fine-tuning a chatbot to engage in conversation with a target open-domain chatbot. The chatbot is fine-tuned with a collection of crafted conversation sequences. Particularly, each conversation begins with a sentence from a crafted prompt sentences dataset. Our extensive evaluation shows that open-domain chatbot models can be triggered to generate toxic responses in a multi-turn conversation. In the best scenario, \toxicbot achieves a 67\% activation rate. The conversation sequences in the fine-tuning stage help trigger the toxicity in a conversation, which allows the attack to bypass two defense methods. Our findings suggest that further research is needed to address chatbot toxicity in a dynamic interactive environment. The proposed \toxicbot can be used by both industry and researchers to develop methods for detecting and mitigating toxic responses in conversational dialogue and improve the robustness of chatbots for end users.

Yuanda Wang, Hanqing Guo, Qiben Yan

Inaudible voice command injection is one of the most threatening attacks towards voice assistants. Existing attacks aim at injecting the attack signals over the air, but they require the access to the authorized user's voice for activating the voice assistants. Moreover, the effectiveness of the attacks can be greatly deteriorated in a noisy environment. In this paper, we explore a new type of channel, the power line side-channel, to launch the inaudible voice command injection. By injecting the audio signals over the power line through a modified charging cable, the attack becomes more resilient against various environmental factors and liveness detection models. Meanwhile, the smartphone audio output can be eavesdropped through the modified cable, enabling a highly-interactive attack. To exploit the power line side-channel, we present GhostTalk, a new hidden voice attack that is capable of injecting and eavesdropping simultaneously. Via a quick modification of the power bank cables, the attackers could launch interactive attacks by remotely making a phone call or capturing private information from the voice assistants. GhostTalk overcomes the challenge of bypassing the speaker verification system by stealthily triggering a switch component to simulate the press button on the headphone. In case when the smartphones are charged by an unaltered standard cable, we discover that it is possible to recover the audio signal from smartphone loudspeakers by monitoring the charging current on the power line. To demonstrate the feasibility, we design GhostTalk-SC, an adaptive eavesdropper system targeting smartphones charged in the public USB ports. To correctly recognize the private information in the audio, GhostTalk-SC carefully extracts audio spectra and integrates a neural network model to classify spoken digits in the speech.