Researchers Manipulate Stolen Data to Corrupt AI Models and Generate Inaccurate Outputs

Researchers propose AURA, a defensive framework that safeguards proprietary knowledge graphs in GraphRAG systems by inserting carefully crafted fake triples (“adulterants”) into critical nodes so that stolen copies produce incorrect outputs while authorized users can filter adulterants using AES-encrypted metadata. The paper details node-selection via Minimum Vertex Cover heuristics, hybrid adulterant generation combining link-prediction models and LLMs, an impact metric (Semantic Deviation Score), and evaluations across multiple datasets and LLMs showing ~94–96% harmfulness (answers flipped wrong) and 100% adulterant retrieval, while noting limitations such as node text handling and insider distillation risks.