Treffer: Optimization and Implementation of Time Series Dimensionality Reduction Anti-fraud Model Integrating PCA and LSTM under the Federated Learning Framework.

For the problem of fraud detection in financial transaction systems, traditional anomaly detection methods based on time series often face challenges such as high computational overhead, inefficient feature extraction, and privacy and security risks when data distribution is highly heterogeneous. To this end, this paper integrates principal component analysis (PCA) and long short-term memory (LSTM) under the framework of Federated Learning (FL) to construct an efficient time series dimensionality reduction anti-fraud model to optimize fraud detection performance and protect data privacy. First, the distributed training of the federated learning framework is adopted to enable different data sources to perform feature extraction and model training locally, thereby achieving data sharing under the premise of protecting privacy. Then, PCA is used to extract the main feature components to improve the effectiveness of data representation. Subsequently, the LSTM model was introduced to perform deep learning on the features after dimensionality reduction to improve the ability to identify abnormal trading patterns. Finally, the Federated Averaging (FedAvg) algorithm is used to aggregate the updates of each local model to optimize the global model performance, while the Federated Differential Privacy (FDP) mechanism is used to enhance data privacy protection. In the experimental conclusion, the proposed model outperforms traditional methods on multiple standard datasets. On the research dataset, the accuracy of the proposed model reaches 93.6%, which is more advantageous than random forest (93.4%) and gradient boosting tree (92.5%). At the same time, the computational overhead and privacy protection capabilities of the model are well balanced. By applying a privacy protection mechanism, although the training time has increased, an effective trade-off between data privacy and performance has been achieved, providing researchers with a new solution. [ABSTRACT FROM AUTHOR]