How to check if dataset has been used for a Neural Network I'm participating in a research by applying some computer vision processing on some images, and I got asked how I could put a watermark on the images in case someone else uses the dataset for purposes not intended on the paper.
Example: Our dataset mustn't be used to create a Convolutional Neural Network with Caffe.
What's the best way to test a dataset has been trained with that data? For example any kind of general test, or applying a watermark in every single image.
 A: This is not a real answer, but rather an idea. 
The reason why I don't immediately dismiss the question as unsolvable: NN problems are usually underdetermined, i.e. num of observations is smaller than num of parameters. In this case, it's very difficult to get rid of noise in parameters. 
Watermarks are basically noise, or not a signal, at least that's how they should appear to a naked eye. So, maybe you can come up with a watermark that will make into parameters of NN. Later once the model is built and deployed, you send a batch of watermarked images into NN, then observe the output. Maybe you can detect that the parameters are contaminated with a watermark by analyzing statistical characteristics of the output images. 
For instance, you'll mark images of dogs and cats with different signatures in the training set. Then for testing NN for theft, you send the same images but with randomly assigned watermarks. If the output shows difference in accuracy of  detection of cats on the same images with different watermarks, it would mean that the NN used your database.
