Custom Binary Cross Entropy Based Anomaly Detection in Bank Transactions using Deep Convolutional Neural Network

A corrigendum was issued for this paper on January 18, 2022. You can download the corrigendum from the supplemental material section of this citation page.

Credit cards have become popular, and their usage has gone up in recent times, due to expansion in e-business and electronic payment systems. The corresponding data is enormous in size and subsequently can cause difficulty in analysis, which has made some organizations to misuse this facility in inappropriate ways, resulting in increase of fraudulent transactions. These transactions are termed as anomalies, whose detection is difficult due to their small number. Anomaly detection with deep learning is a challenging task as the data is highly imbalanced. Vanilla Loss functions usually cannot be used for anomaly detection as certain scenarios require the loss function to assign more loss to false negatives in comparison to false positives. To implement this, we have introduced a hyper parameter, as a modification to the famous loss function Binary Cross Entropy, which can be tuned based on the applications and conditions to get the best results.