Glossary / embedding-inversion

Embedding Inversion

A privacy attack that reconstructs the original input text from a stored embedding vector.

AttackPrivacy

← All terms

Embedding inversion is a privacy attack where the attacker reconstructs the original input text from a stored embedding vector. Modern embedding models lose less information than is intuitive; published research has demonstrated that significant portions of the original text can be recovered with access to the embedding alone and a similar embedding model.

The implication for AI-SPM is that a vector store containing embeddings of sensitive documents is itself a sensitive store, even if the original documents are kept elsewhere. Treating an embedding store as a low-sensitivity index is a recurring audit finding.

Defences include using embedding models with reduced inversion fidelity, applying noise during indexing (with the trade-off of degraded retrieval quality), and treating the vector store with the same access controls as the source documents.

See Embedding Inversion in production.

The Penaxtra platform implements the controls and assessments described above as part of its AI-SPM programme.

AI-SPM platform overview