Network traffic classification is an important part of network monitoring and analysis

which is used for malicious traffic interception

quality of service assurance

application bottleneck prevention and malicious behavior identification.Recently

there are great challenges on the low accuracy and slow convergence problems

when classifying network traffic generated by communication protocols like Modbus and message queuing telemetry transport(MQTT)in the context of power IoT devices applications.To address aforementioned issues

a convolutional neural network-recurrent neural network(CNNRNN)hybrid architecture based on self-attention mechanism is proposed to improve the classification performance of Modbus and MQTT traffic for IoT devices that already in use within the power system.By simulating and collecting the network traffic of power IoT device in the real environment

a large number of Modbus and MQTT communication data packets are obtained.Additionally

the traffic data is converted into pseudo-image format

and a self-attention mechanism is introduced to enhance the network′s attention and feature capture capabilities in different regions.The experimental results show that

compared with traditional multilayer perceptron(MLP)

RNN

CNN models

the proposed CNN-RNN hybrid model with self-attention mechanism demonstrates achieved significant improvement in IoT traffic classification.The model can achieve up to 95%accuracy

demonstrating better convergence and training efficiency.