TY - GEN
T1 - SMART
T2 - 41st Annual ACM Symposium on Applied Computing, SAC 2026
AU - Khozam, Shurok
AU - Blanc, Gregory
AU - Tixeuil, Sebastien
AU - Totel, Eric
AU - Touil, Youssef
N1 - Publisher Copyright:
© 2026 Copyright held by the owner/author(s).
PY - 2026/6/9
Y1 - 2026/6/9
N2 - The proliferation of Internet of Things (IoT) devices within 5G networks has significantly expanded the surface exploited to launch Distributed Denial-of-Service (DDoS) attacks. Due to their limited security capabilities, IoT devices are attractive targets for compromise, enabling adversaries to launch large-scale, high-volume flooding attacks. In Software-Defined Networking (SDN)-based infrastructures, such attacks exploit dynamic, flow-based mechanisms, overwhelming flow tables and congesting forwarding paths. Traditional mitigation strategies struggle to respond to these evolving threats in real-time and typically introduce high latency, making them unsuitable for the performance requirements of modern networks. In this paper, we propose a scalable and generalizable DDoS mitigation framework based on Double Deep Q-Network (DDQN) reinforcement learning algorithm, designed to counter a wide range of flooding-based attacks across multiple protocol layers while preserving Quality of Service (QoS) for legitimate users. Our approach introduces a group-based state projection mechanism to reduce input dimensionality, thereby minimizing computational overhead without sacrificing essential network insights. To ensure adaptability across diverse network topologies, we employ modular neural architectures with permutation-invariant and permutation-equivariant functions, enabling the model to generalize across varying network sizes and entry point configurations without retraining. Experimental evaluations in an emulated SDN environment demonstrate the framework's robustness, scalability, and efficiency in maintaining service availability for legitimate users.
AB - The proliferation of Internet of Things (IoT) devices within 5G networks has significantly expanded the surface exploited to launch Distributed Denial-of-Service (DDoS) attacks. Due to their limited security capabilities, IoT devices are attractive targets for compromise, enabling adversaries to launch large-scale, high-volume flooding attacks. In Software-Defined Networking (SDN)-based infrastructures, such attacks exploit dynamic, flow-based mechanisms, overwhelming flow tables and congesting forwarding paths. Traditional mitigation strategies struggle to respond to these evolving threats in real-time and typically introduce high latency, making them unsuitable for the performance requirements of modern networks. In this paper, we propose a scalable and generalizable DDoS mitigation framework based on Double Deep Q-Network (DDQN) reinforcement learning algorithm, designed to counter a wide range of flooding-based attacks across multiple protocol layers while preserving Quality of Service (QoS) for legitimate users. Our approach introduces a group-based state projection mechanism to reduce input dimensionality, thereby minimizing computational overhead without sacrificing essential network insights. To ensure adaptability across diverse network topologies, we employ modular neural architectures with permutation-invariant and permutation-equivariant functions, enabling the model to generalize across varying network sizes and entry point configurations without retraining. Experimental evaluations in an emulated SDN environment demonstrate the framework's robustness, scalability, and efficiency in maintaining service availability for legitimate users.
KW - deep reinforcement learning
KW - distributed denial of service
KW - double deep Q-Network
KW - scalability
KW - software-defined networking
UR - https://www.scopus.com/pages/publications/105042919362
U2 - 10.1145/3748522.3779843
DO - 10.1145/3748522.3779843
M3 - Conference contribution
AN - SCOPUS:105042919362
T3 - Proceedings of the ACM Symposium on Applied Computing
SP - 1180
EP - 1189
BT - SAC 2026 - 41st Annual ACM Symposium on Applied Computing
PB - Association for Computing Machinery
Y2 - 23 March 2026 through 27 March 2026
ER -