Research On The Application Of Deep Learning In The Detection Of Bank Credit Card Fraud

With the rapid development of internet finance,the issuance of credit cards by various banks in China has also increased at an extremely fast rate every year,which has greatly facilitated transactions between users and merchants,making it fast and convenient.In this context,credit card fraud has also brought serious problems.Fraud will cause great losses to banks and users.Therefore,accurate detection of credit card fraud is an important link to protect the healthy development of the Internet financial industry.In the current field of credit card fraud detection,there are far more non fraudulent transactions in the data set than fraudulent transactions,which is extremely unbalanced and has high dimensions,leading to low classification accuracy and poor prediction performance of existing models.At the same time,the fraud behavior of criminals will change with time,and the characteristics of credit card fraud will also change,leading to overfitting or under fitting problems in model detection,In order to solve the above problems,this research proposes the BERT-CNN-LSTM model to detect credit card fraud and reduce the losses of financial institutions in credit card fraud.The specific research of this paper includes:To solve the problem of extreme imbalance and high dimension in the credit card data set,we preprocessed the data by combining correlation analysis,down sampling and oversampling to balance the data set and various transformations.In the BERT-CNN-LSTM model,we used the pre training of the BERT layer to represent the data as dense vectors,which captured the semantic similarity of the data and used these vectors as input.Specifically,Convert text sequences into token sequences through pre trained word breakers,and learn language representations through maximizing language models and masking language models to form a pre trained model.In this way,the BERT model has the advantage of better understanding the meaning of inputs,providing more accurate detection of fraud and non fraud classification problems,and solving the problem of imbalanced and high-dimensional datasets.The fraud means against criminals will change over time,and the characteristics of credit card fraud will also change,leading to overfitting or under fitting problems in model detection.The BERT-CNN-LSTM model proposed in this article uses CNN layers for feature extraction,concatenates all local feature vectors into a global feature vector through convolutional and pooling layers as input to LSTM,calculates the distribution probability of fraud and non fraud categories through bidirectional hidden layers and fully connected layers,and utilizes the advantage of memory unit components in LSTM to capture long-term dependencies in the sequence,solving the problem of time variation,This leads to overfitting or under fitting of the prediction model.This study uses open data sets to verify the detection performance of the proposed credit card fraud detection model,and sets up ablation experiments and comparative experiments.The experimental results show that the performance of CNN and LSTM models is more advantageous,and compared with traditional machine learning,the model proposed in this paper has stronger advantages,thus a new method for detecting credit card fraud is proposed.