With the widespread popularity and application of internet of things devices， malicious threats in network environments are becoming increasingly frequent. Traditional detection methods are inadequate in dealing with high-dimensional traffic characteristics and multi-scale pattern recognition， and there is an urgent need to build efficient and accurate detection models. Therefore， a malicious traffic detection method combining multi-scale feature extraction and efficient convolution strategy has been proposed. Firstly， introduces the idea of ensemble learning combined with inception structure to construct a network traffic classification model. Secondly， efficient convolutional neural network architectures are adopted to construct a lightweight detection model. Thirdly， channel and spatial attention mechanisms are integrated to enhance feature weighting. Finally， the malicious traffic recognition result is output through the Softmax classifier. The experimental results show that the proposed model is significantly better than the comparative methods in terms of accuracy， precision， and recall. The accuracy reaches 0.94， and the precision and recall are 0.92 and 0.91， respectively. The root mean square error is reduced to 0.07， and the false positive and false negative rates are 0.08 and 0.09， respectively. The average detection delay is controlled at 15 ms， and the parameter count is 6.1 M， demonstrating excellent computational efficiency and stability. The research results indicate that the proposed integrated model can achieve high-precision and low false alarm malicious traffic detection in the internet of things environment， providing reliable support for real-time security protection of intelligent terminals.