Research On Defense Methods Of DDoS Attacks Based On Reinforcement Learning And Federal Learning

Distributed denial of service(DDoS)attacks with very large traffic have become an important threat to network security.The existing DDoS defense methods focus on the end defense.Since the traffic has been pooled,this brings the hidden danger of privacy leakage and service delay to the defense side.At the same time,hybrid DDoS attacks also pose challenges to intermediate network defense methods.Aiming at the shortcomings of the traditional cloud server DDoS filtering method,this paper proposes a method of distributing DDoS defense strategies to edge servers,and using reinforcement learning to optimize defense strategy allocation,and designing a federation learning-based edge server DDoS traffic detection method.The main work is as follows: 1.Propose a graph optimization problem model of network traffic interception and describe its key problems.Set a feasible greedy algorithm for this optimization problem and solve it,then for the defects of this graph optimization problem solution(approximate solution approximation rate,operation time and algorithm scalability),a graph convolution neural network was proposed And deep reinforcement learning solutions,so that it can react to DDoS attack traffic in real time,and change the deployment method of defense strategy.This method uses graph convolutional neural networks to learn the network structure information and vectorize it to provide a unified structure input for the Q network,so that the method can be applied to graph problems of different sizes.2.In view of the current diversified and mixed DDoS attack methods,the protected party's requirements for the security of user information and its own information and the protection of local information by Internet Service Providers(ISP),this paper proposes a classification of DDoS attack traffic based on federal learning The method makes it possible to share learning results on the basis of not exposing local information through edge servers distributed in different ISPs.Experiments show that the reinforcement learning method can effectively reduce the calculation time and obtain an approximate solution.And through experiments,it is proved that the DDoS classifier based on federal learning can make its performance reach the effect of local learning without exposing local information.The method proposed in the article provides new ideas for DDoS attack defense,so that different ISPs can effectively cooperate to defend against attacks and protect their private information.