Abstract: The joint anti-eavesdropping and anti-jamming problem of D2D communications with a probabilistic active eavesdropper is studied in this paper. Since the probabilistic active eavesdropper can launch passive eavesdropping or active jamming with a probability, it is hard to combat. Due to the dynamic change of the active eavesdropper’s attack modes, a robust game learning method is employed to enhance the average secure throughput of D2D communications. We formulate the interaction between a cellular user equipment (CUE) and multiple D2D user equipments (DUEs) to be an one-leader-multiple-follower Stackelberg game. An interference pricing mechanism is introduced to investigate the competitive relationship between the CUE and DUEs, and an exact potential subgame is designed to model the cooperative relationship among DUEs. We have proved that the existence of Nash equilibrium (NE) of the subgame in the lower layer, and further the existence of the Stackelberg equilibrium (SE) of the proposed Stackelberg game. Then a robust and cooperative D2D power control algorithm is proposed to converge to SE, which also outperforms the random and the selfish D2D power control algorithms.