[Paper Review] A Method for Detecting Abnormal Data of Network Nodes Based on Convolutional Neural Network

Yihao Zang, Xianhao Shen+1 more|arXiv (Cornell University)|Jun 28, 2021

Anomaly Detection Techniques and Applications3 references4 citations

TL;DR

This paper proposes a convolutional neural network (CNN)-based method for detecting abnormal data in wireless sensor networks by converting normalized data into grayscale images for input. The approach improves detection accuracy and reliability by learning data characteristics automatically, outperforming traditional threshold-based methods in experimental validation.

ABSTRACT

Abnormal data detection is an important step to ensure the accuracy and reliability of node data in wireless sensor networks. In this paper, a data classification method based on convolutional neural network is proposed to solve the problem of data anomaly detection in wireless sensor networks. First, Normal data and abnormal data generated after injection fault are normalized and mapped to gray image as input data of the convolutional neural network. Then, based on the classical convolution neural network, three new convolutional neural network models are designed by designing the parameters of the convolutional layer and the fully connected layer. This model solves the problem that the performance of traditional detection algorithm is easily affected by relevant threshold through self-learning data characteristics of convolution layer. The experimental results show that this method has better detection performance and higher reliability.

Motivation & Objective

• To address the challenge of unreliable data detection in wireless sensor networks due to faults or anomalies.
• To overcome the limitations of traditional threshold-based detection methods that are sensitive to parameter selection and environmental variations.
• To develop a data-driven approach that learns intrinsic patterns from normal and abnormal data for robust anomaly detection.
• To enhance the reliability and accuracy of data integrity verification in networked sensor systems.

Proposed method

• Normal and abnormal data are normalized and mapped into grayscale images to serve as input for the convolutional neural network.
• Three customized CNN architectures are designed by adjusting convolutional and fully connected layer parameters to optimize feature extraction.
• The network leverages the hierarchical feature learning capability of convolutional layers to automatically capture complex data patterns without relying on predefined thresholds.
• The model is trained end-to-end on image-represented data to classify inputs as normal or abnormal.
• The architecture is based on classical CNN frameworks but modified with tailored hyperparameters for sensor data representation.
• The method enables self-learning of data characteristics, reducing dependency on manual threshold tuning.

Experimental results

Research questions

• RQ1 Can a CNN-based approach effectively detect abnormal data in wireless sensor networks without relying on fixed thresholds?
• RQ2 How does the image-based representation of sensor data improve anomaly detection performance compared to conventional methods?
• RQ3 What is the impact of customized CNN architecture design on detection accuracy and robustness in sensor data applications?
• RQ4 To what extent can the model generalize across different types of injected faults in sensor data?
• RQ5 How does the self-learning capability of the CNN enhance detection reliability in dynamic network environments?

Key findings

• The proposed method achieves higher detection accuracy compared to traditional threshold-based algorithms due to its ability to learn complex data patterns.
• The model demonstrates improved reliability in identifying abnormal data across various fault injection scenarios.
• The use of grayscale image representation effectively preserves the structural characteristics of sensor data for effective classification.
• Customized CNN architectures outperform standard models in terms of detection performance on the given dataset.
• The method reduces sensitivity to parameter tuning, enhancing robustness in real-world deployment.
• Experimental results confirm the model's superior performance in distinguishing normal from abnormal data under diverse conditions.