Towards Quality Controllable Data Synthesis: A Case Study on Synthesizing Network Intrusions

Network intrusion detection plays a significant role in safeguarding network traffic against malicious activities. As the integration of Big Data and Artificial Intelligence (AI) in Intrusion Detection Systems (IDS) becomes more prevalent, achieving high detection accuracy has become increasingly feasible. However, the development of AI-based IDS is often hindered by limited access to diverse and high-quality training datasets, coupled with concerns over sharing raw intrusion data due to privacy and security issues. To mitigate these challenges, a framework that focuses on quality-controllable data synthesis is proposed, specifically for network intrusions. This framework leverages autoencoders for intrusion data generation, ensuring that the synthesized data maintains the characteristics necessary for effective IDS training while enabling controls on the data quality. Our method not only allows for the safe sharing of synthesized intrusion data but also preserves the essential features required for IDS development. We implemented a Deep Learning approach to assess the proposed framework using a benchmark dataset. The results show that our proposed data synthesis approach retains high IDS performance while enabling quality-controllable data synthesis.