Researchers launch AURA to protect AI knowledge graphs

A novel framework called AURA has been unveiled by researchers aiming to safeguard proprietary knowledge graphs in AI systems by deliberately corrupting stolen copies with realistic yet false data.

The approach is designed to preserve full utility for authorised users while rendering illicit copies ineffective instead of relying solely on traditional encryption or watermarking.

AURA works by injecting ‘adulterants’ into critical nodes of knowledge graphs, chosen using advanced algorithms to minimise changes while maximising disruption for unauthorised users.

Tests with GPT-4o, Gemini-2.5, Qwen-2.5, and Llama2-7B showed that 94–96% of correct answers in stolen data were flipped, while authorised access remained unaffected.

The framework protects valuable intellectual property in sectors such as pharmaceuticals and manufacturing, where knowledge graphs power advanced AI applications.

Unlike passive watermarking or offensive poisoning, AURA actively degrades stolen datasets, offering robust security against offline and private-use attacks.

With GraphRAG applications proliferating, major technology firms, including Microsoft, Google, and Alibaba, are evaluating AURA to defend critical AI-driven knowledge.

The system demonstrates how active protection strategies can complement existing security measures, ensuring enterprises maintain control over their data in an AI-driven world.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!