Robust Collaborative Fraudulent Transaction Detection using Federated Learning

Fraudulent transaction detection is a difficult problem for an individual bank, since the number of fraudulent transactions within a single bank’s records is significantly less compared to the day-to-day regular transactions it processes. Hence, due to this extreme data imbalance, training a classifier is difficult. Also, the model will not be able to learn from different types of fraudulent transactions, which a single bank’s database lacks. Collaboration between banks is the only way to achieve a generalized model, but banks will not share their data with each other due to competition and regulatory restrictions. Federated Learning can be leveraged here to solve this problem. However, in a cross-silo setting like this, the data held by different banks will be different in terms of distribution and hence follows a non-IID scenario across the participants’ datasets. Moreover, we are considering that a minority of the banks could be malicious and will try to disrupt this federated learning process. Hence the problem is to perform federated learning in a non-IID setting with active adversaries involved, which is a new research area under fraud detection. We perform non-IID partitioning of the transaction dataset to simulate 10 banks or silos. Then, for benchmark, we perform federated averaging with a subset of the banks set as malicious. Furthermore, we propose a novel algorithm - Epsilon Cluster Selection, a filter-based aggregation technique to recognize and prevent malicious nodes from contributing to the global model being trained. We apply this algorithm to the same setting with malicious banks and compare the results.