A Related Works 318 Private inference has been a promising solution to protect both data and model privacy during deep

Private inference framework CoPriv adopts CypTFlow2 [39] protocol for private inference. MBps bandwidth and 80ms echo latency. We train our proposed CoPriv with self-distillation. For the network re-parameterization mentioned in Section 4.2, here we provide the For some input size, the input cannot be covered by tiles. The correctness and equivalence can be proved with Eq. 1. Also, [ Winograd convolution for stride of 2. The algorithm can be nested with itself to obtain a 2D algorithm The correctness analysis is the same with Section D.3.