A Feature Dimensionality Reduction Method with Lossless Accuracy in Vehicular Network Traffic Detection

In the Internet of vehicles, in-vehicle information systems are protected from malicious attacks by carrying network traffic detection. Existing grayscale-based traffic detection models are not suitable for in-vehicle gateways due to high time and space complexity. Thus, based on feature heat map statistics, this paper proposes a feature dimensionality reduction method with lossless accuracy (LAFDR) for in-vehicle network traffic. The LAFDR solves the problem that the high computational complexity of malicious traffic detection algorithm and the real-time requirements of in-vehicle gateways. The LAFDR uses the Grad-cam algorithm to obtain the feature grayscale figure corresponding to the heat map of network traffic, whose feature weights are counted by the mapping relationship between feature vectors and figure pixel locations. Under the accuracy constraint of the malicious traffic detection model, the method retains the features with high feature weights and eliminates the features with small feature weights. Eventually, on the basis of ensuring the detection accuracy, the LAFDR reduces the computational overhead of the traffic detection model by reducing the input volume of the model. The experiments conducted on the CICDIS2017 dataset, the effectiveness of the LAFDR is illustrated by comparative experiments on the CNN model. The results show that, under the constraint that the detection accuracy of the deep neural network model is no less than 99.0%, the LAFDR achieves the 49.38% reduction in feature dimensionality and the 40% reduction in time overhead.