Modified Higher-Order Dependencies in Networks for Anomaly Detection
Kamala Challa, Alladi Suresh Babu · 2025
Background: Anomaly detection is essential for detecting unusual behaviors in dynamic networks that may represent emerging security threats. Traditional models focus on First-Order Network representations that overlook the integration of higher-order dependencies. Methods: This work proposes a novel Modified Higher-Order Dependencies Network (HON) for anomaly detection. Initially, the input data is preprocessed by min-max normalization, followed by Random Oversampling to address the class imbalance. Then, this preprocessed data is sent to the novel Modified HON that finds the unusual behaviors of anomalies by modeling sequences with n-grams. An Improved Distance Measure is introduced as the loss function combines Improved Euclidean Distance and Interquartile Mean that improves the detection accuracy by prioritizing structural characteristics over node or edge attributes. Result: The proposed approach achieves 94.1% accuracy surpassing existing techniques by up to 12.5%. Conclusion: This technique offers better performance for anomaly detection that paves the way for advanced cybersecurity applications.