A Deep Learning Methodology to Detect Trojaned AI-based DDoS Defend Model

DDoS attack arranges bots to send low-speed traffic to backbone links and paralyze all servers in the target area. DDoS is difficultly defended due to the two research problems (1) indistinguishability of the changing DDoS characteristics and (2) the time series attack pattern, leading that the raising attention of developing varying DDoS defending methodologies. The conventional methods to defend DDoS apply a rules-based methodology that relies on the experience of algorithm designers and cannot reflect the changing attack characteristics of DDoS in a timely manner. Numerous artificial intelligence (AI) methodologies, therefore, are introduced to defend DDoS through end-to-end functionality (Input: network status; Output: defending action) without any manual intervention. However, the AI-based DDoS Defending model often outsources training to a machine-learning-as-a-service (MLaaS) provider because of the scarce training dataset and high hardware requirement. This may cause the model been trained maliciously, which is called the Artificial Intelligence Trojan attack (AI Trojan). AI Trojan means an AI model encounters a malicious training process and then have a good performance on normal data but behaves maliciously with certain data. This study proposes GAN based AI Robustness test algorithm, Deep Learning Attack Generator (DLAG), to verify that the artificial intelligence model has been fully trained to ensure the robustness of the model. DLAG can be divided into five steps: (1) DLAG randomly generates a sample, (2) generates noise that participates in the generation of a confrontation network (DLAG), (3) input the synthetic sample to the testing AI, (4) the test results will be recorded in the test report and fed back to GAN, and (5) a new synthetic sample will be generated again for the next test cycle. The simulation shows that our proposed DLAG can detect that the AI based DDoS/LFA detector is trained by imbalance data. The simulation results also demonstrate the potential and suggested development trait of AI Trojan detection methodology.