Latency vs. Privacy: A Theoretical Architecture for Pre-Warmed Trusted Execution Environments in Real-Time Bidding

Abstract

Real-Time Bidding has a high set of temporal limitations since advertising platforms have to be fast and capable of responding within about 100 milliseconds to compete effectively in online auctions. With the development of privacy-first models, which require Trusted Execution Environments, there grows a basic system incompatibility, with standard instantiations of these secure computational models, where the size of the initialization sequences is orders of magnitude larger than the time required to run a single auction. The cold start phenomenon, including secure boot sequence and remote attestation protocol implementations, causes latencies that make on-demand provisioning incompatible with real-time bidding needs. The Pre-Warmed Enclave Pool architecture mitigates this problem by keeping a persistent collection of fully initialized and attested secure enclaves, practically separating the costly trust establishment steps in the context of a request and its important processing path. Projections based on mathematical modeling of the performance of recorded attributes of a commercial trusted execution platform show that this architectural style is capable of decreasing end-to-end processing latency to the point of achieving the required thresholds of auction participation. The resolution requires one to accept more infrastructure spending, which is a privacy tax, which is the costs that are necessitated by the need to over-provide resources to accommodate variable traffic patterns whilst maintaining sufficient service levels. According to this architectural blueprint, it is possible to combine privacy compliance with performance optimization, but in order to achieve both goals at the same cost, organizations must bear increased computational expenses to fulfill regulatory compliance and protect user information as the economic cost of regulatory compliance and user-data privacy.