Vector Store Encryption

Protects vector data stored on disk using industry-standard algorithms.

• AES-256: The Advanced Encryption Standard with a 256-bit key is the default for securing database files and indexes.
• Transparent Data Encryption (TDE): Encrypts the database at the file level, requiring no changes to the application logic.
• Key Management: Integrates with external Key Management Services (KMS) like AWS KMS or HashiCorp Vault to securely store and rotate encryption keys.

Secures vector data as it moves between clients, application servers, and the database.

• TLS 1.3: Enforces the latest Transport Layer Security protocol for all client-server and inter-node communication.
• mTLS: Mutual TLS authenticates both the client and the server using certificates, preventing man-in-the-middle attacks.
• gRPC Security: For databases using gRPC, all remote procedure calls are encrypted by default using the HTTP/2 protocol with TLS.

Encrypts embeddings within the application before they ever reach the database, ensuring the storage layer never handles plaintext data.

• Client-Side Field-Level Encryption: The application encrypts specific vector fields using a master key unknown to the database server.
• Bring Your Own Key (BYOK): The enterprise retains sole control of the root encryption key, making data unreadable to the cloud provider.
• Zero-Knowledge Architectures: The service provider has no technical ability to decrypt customer vector data, as keys are managed exclusively on the client side.

A privacy-preserving technique that allows similarity searches to be performed directly on encrypted vectors without ever decrypting them.

• Computation on Ciphertext: The database performs mathematical operations on encrypted data, returning an encrypted result that only the client can decrypt.
• Approximate Nearest Neighbor (ANN) over Encrypted Data: Specialized algorithms enable efficient search over ciphertext, though with higher computational overhead.
• Use Case: Ideal for highly regulated industries where even the database administrator must not see query vectors or results.

Automated processes for managing encryption keys to limit the impact of a potential key compromise.

• Periodic Rotation: Keys are automatically rotated at defined intervals (e.g., every 90 days) without downtime.
• Data Re-encryption: The process of decrypting data with an old key and re-encrypting it with a new one, often performed as a background task.
• Key Versioning: Maintains a history of keys to allow decryption of older data that was encrypted before the latest rotation.

Leverages dedicated, tamper-resistant hardware to manage and safeguard cryptographic keys.

• FIPS 140-2 Level 3: HSMs certified to this standard provide strong physical and logical protection for key material.
• Offloading Crypto Operations: The HSM performs all cryptographic operations, ensuring private keys never leave the secure hardware boundary.
• Centralized Key Trust: Establishes a hardware root of trust for the entire vector store encryption infrastructure.

A privacy-preserving computation method that allows similarity searches to be performed directly on encrypted vectors without ever decrypting the underlying data. The database processes mathematical operations on ciphertexts, returning encrypted results that only the client can decrypt. This ensures that even the database administrator or a compromised server cannot view the raw embeddings or the query itself. Partially homomorphic encryption (PHE) is often used for distance calculations, while fully homomorphic encryption (FHE) remains computationally expensive for production workloads.

The process of applying a reversible or irreversible transformation to a vector to mask its true semantic meaning from unauthorized observers. Techniques include:

• Random rotation: Applying a secret orthogonal matrix to rotate embeddings while preserving cosine similarity.
• Dimensionality reduction noise: Projecting to a lower dimension and adding calibrated noise.
• One-way hashing: Irreversibly transforming embeddings for indexing without revealing original features. This allows third-party services to perform useful operations on vectors without accessing the raw semantic content.

A data protection technique that adds random mathematical noise to embeddings to degrade the accuracy of unauthorized similarity searches while preserving utility for legitimate queries. The noise is calibrated to the sensitivity of the data and the desired privacy budget (ε). Legitimate users with the correct decryption keys or noise-cancellation parameters can still retrieve accurate results. This technique is closely related to differential privacy and is effective against model inversion attacks and membership inference attempts.

A data isolation strategy that applies distinct encryption keys and access policies to individual physical or logical shards of a vector index. Each partition can be encrypted with its own AES-256-GCM key, managed by a Key Management Service (KMS). This ensures that a breach of one partition does not expose the entire vector store. Partition-level security is essential for multi-tenant architectures where different clients' embeddings must remain cryptographically isolated even within the same database cluster.

A cryptographic protocol that allows multiple parties to jointly perform a vector search across their private datasets without revealing their individual data to one another. Using secure multi-party computation (SMPC) or private set intersection (PSI) techniques, each party encrypts their embeddings and participates in a distributed computation. The final result reveals only the top-k nearest neighbors that match the query, without exposing the underlying vectors. This is critical for federated search across competing organizations or regulated data silos.

The real-time monitoring of vector database egress traffic to identify and block unauthorized attempts to transfer large volumes of embeddings or associated metadata. Detection systems analyze:

• Query volume anomalies: Unusual spikes in the number of vectors returned per query.
• Semantic drift: Queries that systematically walk through the embedding space to reconstruct data.
• Temporal patterns: Automated scraping behaviors at regular intervals. When combined with semantic rate limiting and query fingerprinting, this forms a defense-in-depth strategy against extraction attacks.