Hybridi autoenkooderi-random forest-malli palvelunestohyökkäysten tunnistamiseksi verkkoliikennelokidatasta
Year: 2025
Type: dissertation
Abstract: This master’s thesis explores whether combining unsupervised deep learning based autoencoders with classical ensemble methods can yield practical gains in detecting distributed denial-of-service attacks from network log data. Random forests have already been established as robust, high-accuracy tools for the task by prior research but they utilize shallow learning and may overlook deeper non-linear patterns in complex data. As attacks grow more sophisticated, models capable of capturing deeper patterns and relationships in the data are increasingly necessary to maintain detection performance in the evolving environments. Two detection workflows were built to test the premise: a baseline random forest and a hybrid autoencoder-random forest model. Both models were trained and evaluated using the publicly available and widely used CICDDoS2019 dataset. In the hybrid model, the deep-learning-based autoencoder compresses the data into a 16-dimensional latent representation and feeds the most relevant features based on reconstruction error to the random forest for classification. The results show that the hybrid model was able to match the baseline’s high detection rates of around 99% accuracy with comparable precision, recall, F1-score and ROC-curves, matching the expected results of current benchmarks in research. The hybrid model cut the training time of the random forest model by more than half, though inference times increased by 77%. The findings suggest adopting the hybrid approach in slower batch-processed scenarios or forensics-like environments where the deep learning’s benefits are best leveraged. The standalone random forest remains advantageous in latency-critical deployments in high-volume real-time conditions.
Source: Aaltodoc (Aalto University)
Author: Kunnas, Juho
Language: English
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W6991581884
