Evaluating Reinforcement Learning Models for Optimized Runtime API Threat Detection

The increasing use of application programming interfaces has revolutionized the world of modern software systems, providing scalability in cloud, mobile, and microservices ecosystems. However, this rise has also resulted in a broad attack surface and a considerable rise in cyber threats associated with the misuse of APIs during their runtime. Traditional detection systems have remained largely rule-based on static characteristics and have relied on supervised learning techniques. However, they lack flexibility and have failed to keep up with the dynamically changing attack patterns. The current research focuses on the use of reinforcement learning models to improve the detection capabilities of APIs. The problem statement represented using APIs has been tested using a variety of reinforcement learning techniques, and the results indicate that agents developed using these models have better adaptability and resistance to novel attack patterns. However, trade-offs have been found between detection accuracy and latency.