SGANFuzz: A Deep Learning-Based MQTT Fuzzing Method Using Generative Adversarial Networks
Zhiqiang Wei, Xijia Wei, Xinghua Zhao, Zongtang Hu, Xu Chu
Abstract
As the Internet of Things (IoT) industry grows, the risk of network protocol security threats has also increased. One protocol that has come under scrutiny for its security vulnerabilities is MQTT (Message Queuing Telemetry Transport), which is widely used. To address this issue, an automated execution program called fuzz has been developed to verify the security ofMQTTbrokers. This program is provided with various random and unexpected input data and monitored for different responses, such as acknowledgments, crashes, failures, or memory leaks. To generate a significant number of realistic MQTT protocols, we have proposed a Generative Adversarial Networks (GAN)-based protocol fuzzer called SGANFuzz. Our experimental results showthat SGANFuzz has successfully detected 6 vulnerabilities among 7MQTTimplementations, including 3 CVE bugs. Compared to the state-of-the-art fuzzing tools, SGANFuzz has proven to be the most efficient fuzzing tool in terms of vulnerability detection and has expanded the feedback coverage by receiving more unique network responses from MQTT brokers.