This project is a simple PyTorch implementation of "One pixel attack for fooling deep neural networks" on the Cifar10 dataset. The code is developed upon Pytorch-cifar and one-pixel-attack-keras.
| model | Accuracy on the test set | Success Rate (1 pixel, untargeted) | Success Rate (3 pixels, untargeted) |
|---|---|---|---|
| vgg16 | 93.42% | ~50.0% | ~93.0% |
| res18 | 94.94% | ~27.7% | ~78.0% |
| res101 | 94.51% | ~19.0% | ~63.3% |
To speed up working with this code, we've included copies of parts of the PyTorch_CIFAR10 repository published by Huy Phan. This project enables us to easily use pretrained models and avoid the expense of retraining.