How to Make a Convolution Kernel in a Layer Not Participate in Back Propagation in Convolutional Networks?

I’m doing research on model pruning.
Filter pruning.(Kernel pruning)
I set the value of some filters to zero
then retrain model. But these values become non-zero in the process of Back Propagation.

What can I do to solve this problem? (Keep these filters at 0 during retraining)

Thanks for your help~~