Anomaly Detection in Provenance Data Using Autoencoder Network

This work aims at analyzing how provenance data can be used in anomaly detection by employing autoencoder networks which is a crucial operation for securing the various sectors through validation of data derived from cybersecurity, research, among other fields. The experimental data samples which have been used include ProvBench, DARPA OpTC, and W3C PROV Challenge datasets used throughout the research process the proposed methodology encompass extensive data pre-processing where the quality of data was normalized and its dimensions reduced through practices such as PCA then an autoencoder was trained to detect the anomalies based on its reconstruction errors. In detailed evaluation on the ProvBench dataset, the model achieved high accuracy of 0. 91, recall of 0. 87, F1-score of 0. 89, and ROC-AUC of 0. 92, much higher than those of the traditional algorithm and non-structural models including the Support Vector Machine and Random Forest. This present study demonstrates the effectiveness of the autoencoder in carrying out generalization across various datasets while being robust and relatively fast in computation. It was also revealed that though there is the need for a cautious selection of the threshold as well as hyperparameter of the autoencoder, the technique of anomaly detection enables high effectiveness in the proved provenance data sets. More works need to be done in improving its adaptability as well as exploring the use of hybrid models of detection in order to provide better results. These presented possibilities provide a robust strategy for ensuring the integration and protection of provenance data in different applications.