This project tests Neural Networks' robustness against Single Event Upset.
A pretrained Neural Networks, invert one of its weights' first bit, test the disrupted network on a validation set, then record the accuracy loss. Test every bits (or randomly sample some bits) of this network, analyze the effect of their inversion.
Experiments are conducted mostly on XNOR-Net
Used an PyTorch implementation of the XNOR-Net. Major networks are as follows:
| Dataset | Network | Accuracy |
|---|---|---|
| MNIST | LeNet-5 | 99.23% |
| CIFAR-10 | Network-in-Network (NIN) | 86.28% |
| ImageNet | AlexNet | Top-1: 44.87% Top-5: 69.70% |