Privacy-Preserving ANN Search

Oblivious approximate nearest neighbor search within Trusted Execution Environments (Under Review)

Privacy-Preserving Approximate Nearest Neighbor SearchUnder Review

Problem

Running approximate nearest neighbor (ANN) search inside Trusted Execution Environments (TEEs) is attractive for privacy-sensitive vector workloads, but graph-based indexes leak information through memory access patterns, exposing the system to side-channel attacks.

Approach

We propose a secure ANN index structure within TEEs that mitigates side-channel attacks by enforcing oblivious memory access patterns on proximity graph indexes. A novel distribution-aware oblivious access algorithm is used to retain plaintext-level throughput while providing provable privacy guarantees.

Results

  • Plaintext-level throughput with provable obliviousness.
  • Significantly outperforms SOTA oblivious ANN solutions.