How to Mislead AI-Assisted Network Automation in SD-IPoEONs: A Comparison Study of DRL- and GAN-Based Approaches

Recently, the combination of artificial intelligence (AI) and software-defined networking (SDN) has attracted intensive research interests because it realizes and promotes AI-assisted network automation (AIaNA). Despite the initial successes of AIaNA, its vulnerabilities, i.e., the downside of the reduction of human involvement achieved by it, have not been carefully explored. In this work, we use software-defined IP over elastic optical networks (SD-IPoEONs) as the background, and study how to mislead the AIaNA system in them. Specifically, we target our attack on the deep neural network (DNN) based traffic predictor in the AIaNA system, and design an adversarial module (ADVM) that can craft and inject adversarial traffic samples adaptively to disturb its operation. We consider two approaches to design the ADVM, i.e., the deep reinforcement learning (DRL) based on deep deterministic policy gradient (DDPG), and the generative adversarial network (GAN) model. Our proposed ADVM can monitor and interact with a dynamic SD-IPoEON to train itself on-the-fly. This enables it to generate and inject adversarial samples in the most disturbing and hard-to-detect way and to severely affect the AIaNA's performance on multilayer service provisioning. Specifically, IP flows will be served incorrectly to result in unnecessary congestions/under-utilizations on lightpaths, and erroneous network reconfigurations will be invoked frequently. Simulation results confirm the effectiveness of our ADVM designs, and show that the GAN-based ADVM achieves better attack effects with smaller perturbation strength.