Privacy-Preserving Anomaly Detection in Encrypted Traffic

Sammanfattning

In 2023, 90% of internet traffic was encrypted which blocked forensic analysis without opening the packets. Our thesis develops a method for the identification of suspicious behavior in encrypted traffic on the basis of metadata, without breaking privacy laws including GDPR. Four machine learning algorithms were experimented with for anomaly detection: Isolation Forest, One-Class SVM, DBSCAN, and K-Means. Based on the results, Isolation Forest was selected for the final system due to its superior performance (AUC = 0.94, AP = 0.36). We have developed thirteen visualizations including scatter plots and performance charts, to clearly show the results, and finally with the help of seven peers in network forensics we reviewed these plots to find the better solution. Our system can be used in Security Operations Centers (SOC) and it will be helpful in investigating threats such as data theft while protecting privacy. Our system will provide accuracy, clarity and legal compliance. Keywords: Encrypted Traffic, Digital Forensics, Machine Learning, Anomaly Detection, PCA, DBSCAN, Isolation Forest, Interactive Visualization