Our code using Flower
and @pytorch.org
is available at: github.com/ellisalicant... !

while in an IID setting, we observe better client privacy with the same level of global model performance, our results show that for non-IID data, the model integration matters, and for privacy-conscious clients opting for smaller models can be counterproductive

we adapt the premise of "bigger models overfit the data" to heterogeneous federated learning to investigate if using different model complexity in clients can lead to better privacy against membership inference attacks