Partition-Level Security

Enforces access control at the granularity of individual vector embeddings. This mechanism ensures that a similarity search only returns results from vectors the user is explicitly permitted to see, preventing cross-contamination of sensitive data within a single collection.

• Granularity: Operates on single embeddings, not entire files.
• Mechanism: Applies ACLs during the Approximate Nearest Neighbor (ANN) search.
• Use Case: A multi-tenant knowledge base where different clients' documents are stored in the same collection.

A pre- or post-query access control technique that restricts vector search results by applying Boolean constraints on associated document tags, timestamps, or user permissions. This is often the first line of defense before a semantic search is executed.

• Pre-filtering: Narrows the search space before ANN, improving performance and security.
• Post-filtering: Applies rules to the final result set, ensuring no unauthorized vectors slip through.
• Example: user_role == 'engineer' AND clearance_level >= 3

A multi-tenancy architecture that logically or physically partitions vector indexes to ensure strict data isolation between different organizations or business units. This prevents any possibility of a query from one tenant accessing another's data.

• Physical Partitioning: Each tenant gets a dedicated index or collection.
• Logical Partitioning: A single index uses a partition key to enforce isolation.
• Performance Impact: Physical isolation can increase resource overhead but provides the strongest security guarantee.

Defensive techniques used to prevent adversaries from reconstructing sensitive source data by analyzing model outputs. These attacks exploit the model's memorization of training data.

• Differential Privacy: Adds calibrated noise during training to provide a mathematical guarantee against data reconstruction.
• Output Perturbation: Adds noise to query responses to obscure fine-grained details.
• Rate Limiting: Restricts the number of queries an entity can make to prevent the assembly of a complete dataset from partial outputs.

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 the query and returns an encrypted result that only the client can decrypt.

• Fully Homomorphic Encryption (FHE): Supports arbitrary computations on ciphertexts but is computationally intensive.
• Practical Application: Allows a cloud-hosted vector database to serve queries without ever seeing the user's query vector or the stored data in plaintext.

A throttling mechanism that restricts the number of vector queries a user can make based on the conceptual topic of the query, not just the raw request count. This prevents automated data scraping that pivots through semantic space.

• Topic Clustering: Groups incoming queries by their semantic meaning.
• Dynamic Thresholds: Applies stricter limits to queries targeting sensitive conceptual clusters (e.g., 'financial projections', 'PII').
• Goal: Detect and block an attacker who is systematically probing a knowledge base with semantically similar but syntactically different queries.