Researchers have developed a new approach to AI security that employs text prompts to better protect AI systems from cyber threats. This method focuses on the creation of adversarial examples to prevent AI from being misled by inputs that are typically undetectable to humans.
The prompt-based technique streamlines the generation of these adversarial inputs, allowing for quicker response to potential threats without extensive computations. Preliminary testing has shown that this method can effectively safeguard AI responses with minimal direct interaction with the AI systems.
Dr. Feifei Ma, the lead researcher, outlines the process: “Our approach involved initially crafting malicious prompts to identify vulnerabilities in AI models. Following this identification, these prompts were utilized as training data, helping the AI to resist similar attacks in the future.”
Subsequent experiments indicated that this training approach improved the robustness of AI systems. Models trained with adversarial prompts were less likely to succumb to similar attacks, demonstrating an enhancement in their defensive capabilities.
“This method allows us to expose and then mitigate vulnerabilities in AI models, which is especially critical in sectors like finance and healthcare,” Dr. Ma noted.
The research indicates that AI systems trained with these adversarial prompts are more capable of resisting similar manipulation tactics in the future, potentially improving their overall robustness against cyber threats.
This research has been published in Frontiers of Computer Science and is a collaborative work between Chinese Academy of Sciences, University of Chinese Academy of Sciences, Stanford University, and National University of Singapore. The complete study is accessible via DOI: 10.1007/s11704-023-2639-2.
