Combine with multiplexed convolution

[Destinyluke]

I recently had the opportunity to read your paper "SpENCNN: Orchestrating Encoding and Sparsity for Fast Homomorphically Encrypted Neural Network Inference". In the paper you mentioned that for convolution with stride > 1, you used Lee's multiplexed convolution for the ResNet-20. Since the CKKS part of your code is not completed, could you please help me understand how to combine multiplexed convolution in detail?

[ranran0523]

When you have group-convolution layers, you can skip the step for summation of all input channels shown in Figure 4 (b) of Lee et al ICML2022. Further, you can also use sub-block pruning to reduce the number of required rotations of multiplication with filter coefficients in Figure 4 (a).

[Destinyluke]

Thanks for your reply. However, I still find it hard to understand how to skip the step for summation of all input channels shown in Figure 4(b). Could you please provide a figure similar to Figure 4 of Lee et al ICLM2022? I greatly appreciate your assistance.