A Black-Box Attack Method against Machine-Learning-Based Anomaly Network Flow Detection Models

Figure 2

Abnormal network flow adversarial example generate process. We (1) choose an initial network flow training set and (2) generate subdataset by sampling algorithm, then (3) label the by the target DNN model and (4) train the substitute DNN model , finally (5) craft the adversarial example with the substitute DNN model , the abnormal network flow data , and the physical limitations on network flow .