Anomaly prediction of Internet behavior based on generative adversarial networks

Area of Science:

• Computer Science
• Artificial Intelligence
• Cybersecurity

Background:

• Internet applications generate vast amounts of user behavior data.
• Abnormal employee internet activities pose significant risks to information security and data integrity.
• Manual labeling of big data for anomaly detection is costly and inefficient.

Purpose of the Study:

• To propose an unsupervised generative model, APIBGAN, for effective anomaly prediction of internet behaviors.
• To address the challenge of high labeling costs in big data anomaly detection.
• To develop a robust method for identifying abnormal employee internet activities.

Main Methods:

• Developed Anomaly Prediction of Internet behavior based on Generative Adversarial Networks (APIBGAN).
• Utilized a data-generating generative adversarial network (DGGAN) to learn real data distribution and generate labeled data.
• Employed a distance-based anomaly prediction approach using generated data as a benchmark.
• Trained APIBGAN on three categories of corporate employee internet behavior data.

Main Results:

• APIBGAN achieved prediction scores of 87.23%, 85.13%, and 83.47% on different datasets.
• These scores surpassed the 81.35% achieved by the Isolation Forests comparison method.
• Experimental results validated APIBGAN's effectiveness in predicting internet behavior outliers using Generative Adversarial Networks (GANs).

Conclusions:

• APIBGAN effectively predicts internet behavior anomalies using a GAN composed of simple fully connected neural networks.
• The model demonstrates broad applicability for anomaly prediction in scenarios with large, difficult-to-label datasets.
• This research offers valuable insights into GAN-based anomaly prediction techniques.