Deployed in a harsh or hostile environment and running in an unattended state, wireless sensor networks (WSNs) are vulnerable to many attacks. The selective forwarding attack is one most difficult to be detected, and trust-based detection methods often fail especially under the poor wireless channels. Based on nodes' cumulative forwarding rates (CFRs) and cumulative transmission rates (CTRs), we propose a clustering scheme in clustered WSNs to divide the nodes into three types: malicious, suspicious, and regular nodes. To defend selective forwarding attacks of malicious nodes, our scheme isolates them from the network. For suspicious nodes, our scheme uses the non-cooperative game with incomplete information to force them to promote forwarding rates. In the game, the reward and punishment mechanism reduces the attackers' expected revenues and trust values. And the suspicious nodes, screened out by our clustering method, will be forced to forward packets to elude our detecting scheme. The Nash equilibrium of the game between regular nodes and suspicious nodes is proved to exist here. Simulation results show that our scheme can promote network throughput largely, and our dynamic-time behavior monitoring scheme can gain a longer network lifetime than that of the full-time behavior monitoring scheme.INDEX TERMS Cluster-based wireless sensor networks, density-based clustering, game theory, network throughput, Nash equilibrium, selective forwarding attack.