Advancing IoT Anomalous Traffic Detection through Arithmetic Coding in Transformer-Based Architectures

With the widespread proliferation of mobile devices and the rapid development of mobile network technologies, the issue of network security has become increasingly critical, especially in the field of network anomaly detection. Traditional anomaly detection methods, such as rule-based and feature-matching techniques, often struggle to effectively cope with increasingly complex network threats. To address this issue, we have incorporated machine learning and deep learning technologies, particularly a strategy of integrating representative tags into word vectors through arithmetic coding. This method, inspired by position encoding, directly enhances the expressive power of the original word vectors and improves the model’s precision in recognizing anomalous behaviors in industrial IoT traffic data, such as ransomware, crypto-mining malware, and DDoS Trojans. This encoding strategy uses shorter codes for frequently occurring words and longer codes for less common words, which not only makes the encoding of information more efficient but also helps the model learn and adapt to the data’s characteristics more quickly. Our experiments have demonstrated that this approach significantly improves fitting efficiency and detection accuracy compared to traditional models without such encoding. This indicates that combining machine learning technologies with innovative data processing methods can significantly optimize the performance of network anomaly detection.