Because of GDPR and confidentiality issues, no bank will expose the user's data to third-party researchers, and hence we cannot train our model for real-time fraud detection with real data (What I mean by real data is, that the data has not been modified | changed by any mean). All researchers have worked on a common data set, which is 7 years old, and that might not have all the Features that potentially changed over time. So, I want to use the federate learning Algorithm to train the model with real dataset to follow GDPR law and access to information while preserving the privacy of users.
This seems to be very practical, code-oriented tutorial on federated learning with tensorflow here:
https://towardsdatascience.com/federated-learning-a-step-by-step-implementation-in-tensorflow-aac568283399
It does not directly apply to fraud detection though.
Also, there are a number of anonymized datasets of credit fraud that seem to be more updated than the 2013 data set you mention, such as here:
https://www.kaggle.com/c/ieee-fraud-detection
as well as a number of other ones on Kaggle.com. That would seem to be one way to deal with privacy concerns and the older dataset you mentioned.
Following tutorial uses the classic MNIST dataset to introduce the federated learning for image classification:
https://www.tensorflow.org/federated/tutorials/federated_learning_for_image_classification
- Badges
- Science topic