MNIST is too simple to make a compelling paper on its own. Model inversion attack works quite well because images are so simple. We need to use a more "real-world" dataset to present a compelling narrative.

Possible datasets:

• CIFAR10/100
• SVHN (driverless cars are leaking our house information!)
• COCO (can we make out human faces?)

Apply:

• nopeek
• DP
• network architecture experiments