Is there any documentation on the FHE-specific implementation choices made for each workload?
For example, how ReLU is approximated in ResNet-20, or what exact architecture is used for the CIFAR-10 benchmark?
These choices significantly affect performance in FHE, so they're an important context for anyone trying to reproduce or compare results.
Is there any documentation on the FHE-specific implementation choices made for each workload?
For example, how ReLU is approximated in ResNet-20, or what exact architecture is used for the CIFAR-10 benchmark?
These choices significantly affect performance in FHE, so they're an important context for anyone trying to reproduce or compare results.