Security-Communication-Computation Tradeoff of Split Decisions for Edge Intelligence

Security and privacy are of paramount importance to data-driven edge intelligence services. By offloading computation-intensive model portions to the edge server, split inference can empower low-latency and energy-efficient model inference at resource-constrained devices. The split decision is critical to the communication and computation performances of split inference. However, its impact on security issues has yet to be studied. This article first investigates the inference security-communication-computation tradeoff when split at different model layers. With the emphasis on security concerns, we summarize the threat models of split inference; and we illustrate two passive attack mechanisms for recovering the input data and private labels. We show case studies to validate the security-communication-computation tradeoff; and we also investigate optimal split-point selection, according to the specific device capabilities and service requirements.