This is a very bad idea.
No matter how good the algorithm, there's always going to be a false positive. In this case, false positive means someone with legitimate account is going to be unable to use the currency, or is going to have great problems with it.
False positives are acceptable in centralized systems, because if false positives are sufficiently rare, there's always some authority who can manually allow someone. No such thing here.
Also, it's not a problem now. No point wasting time fighting an imagined problem, there's always something real to improve.
If you absolutely must execute this bad idea, don't use neural networks. They're the most known, but very hard to get right, they're very prone to overfitting, and choosing the right structure is often pure guesswork.
Use random forests. They're drastically easier to train, prediction is much faster and completely parallel, and results are very often much better than those from neural networks.
For extra accuracy, use also SVM (support vector machine) and combine the results of the two classifiers into something called an ensemble classifier. Ensembles made of many algorithms are the current state of the art.
