Ghost Vectors: Soft-Deleted Embeddings Remain Reconstructible in HNSW Vector Databases

The paper discusses vulnerabilities in HNSW vector databases regarding soft-deleted embeddings, which remain recoverable despite deletion requests, raising compliance issues with data protection regulations like GDPR and HIPAA. The authors demonstrate that using the Vec2Text inversion model, they can recover significant amounts of sensitive data from real-world datasets, achieving 100% recovery for structured data like patient demographics. To address this, they propose Epoch Key Rotation, an encryption method that ensures deleted vectors are unrecoverable, completing the process in 2.5 ms for 500 records while providing cryptographic proof of deletion, which is critical for practitioners concerned with data privacy and compliance in AI applications.