Signature-Based Traffic Classification and Mitigation for DDoS Attacks Using Programmable Network Data Planes

Distributed Denial of Service (DDoS) attacks mitigation typically relies on source IP-based filtering rules; these may present scaling issues due to the vast amount of involved sources. By contrast, we propose a source IP-agnostic DDoS traffic classification and filtering schema that identifies malicious packet signatures via supervised Machine Learning methods and subsequently generates signature-based filtering rules. To accelerate packet processing, our schema utilizes XDP middleboxes operating as programmable Deep Packet Inspectors. Signatures are extracted from network traffic as unique combinations of the most significant packet features; these are subsequently fed to supervised Machine Learning algorithms that classify them as malicious or benign. Malicious signatures undergo a reduction process tailored to the attack vector in order to generate a concise set of filtering rules, thus expediting mitigation performance. Our schema was implemented as a proof-of-concept and evaluated for DNS volumetric attacks in terms of signature classification accuracy and packet filtering throughput. Experiments were based on benign and malicious traffic datasets recorded in production network environments. Our approach was compared to source-based mechanisms in terms of (i) malicious traffic identification, (ii) filtering rules cardinality, and (iii) packet processing throughput required in modern high speed networks. The experimental results demonstrate that our signature-based approach outperforms IP-based alternatives, achieving high detection accuracy and significant generalization capabilities.