Onyx: Cost-Efficient Disk-Oblivious ANN Search

Onyx is a newly developed system that tackles the problem of conducting efficient oblivious approximate nearest neighbor (ANN) searches on SSDs within trusted execution environments (TEEs). Current ORAM-ANN solutions focus on reducing access counts at the ANN level while minimizing bandwidth at the ORAM level, often resulting in an imbalance that overuses both resources. In contrast, Onyx reverses this approach by prioritizing bandwidth reduction in the ANN layer and access count minimization in the ORAM layer, leading to enhanced balance and cost-effectiveness. This system is intended for AI applications managing sensitive information on third-party platforms, where external SSDs may expose user queries through their access patterns. Onyx seeks to lower latency and enhance cost-efficiency relative to existing leading methods.