This paper studies an optimal transmit power decision policy for energy-efficient data transmissions between a sensor node (i.e., the source) and a cluster head (i.e., the destination) in cluster-based wireless sensor networks in the presence of a full-duplex (FD) active eavesdropper. In this network, the source is powered by a wireless energy harvester, while the destination is constantly supplied by the traditional electrical energy. The eavesdropper is capable of FD transmitting and receiving and hence opportunistically launches jamming attacks against the destination while eavesdropping, which affects not just the legitimate transmissions but the eavesdropper itself. The destination can also work in the FD mode to simultaneously receive information signals and send an artificial noise to interfere with the eavesdropper. Therefore, we investigate an optimal power allocation policy for the source in order to maximize the secrecy transmission rate against an FD eavesdropper. In addition, we study the problem of decision making in two different scenarios. First, the legitimate nodes are assumed to have prior information about the arrival of harvested energy and about the eavesdropper's jamming attack model. The problem is formulated as the framework of a partially observable Markov decision process and is solved with value iteration-based dynamic programming. Second, the legitimate nodes do not know the dynamics of the environment in advance, so the problem becomes a standard Markov decision process. Hence, we propose an actor-critic learning framework to find the solution from practical interactions with the environment. Finally, we verify the performance of the proposed schemes by simulations.