In cognitive radio system, the secondary users learn radio spectrum environment through spectrum sensing to get the spectrum holes that the primary users do not occupy. In practice, the existence of various malicious attacks can seriously affect the reliability of spectrum sensing of the secondary users. Only in-depth study of the malicious attack strategies can ensure the security of cognitive radio networks. Based on this, a spoofing jamming strategy in cognitive wireless network, called as primary user emulation attack strategy, was studied. The strategy deteriorates the spectrum sensing performance of the secondary users by transmitting the forged primary user signals over channels. More concretely, the attack strategy was modeled as an online learning problem, and a Thompson sampling based attack strategy was proposed to find an efficient tradeoff between the exploitation of high-performance channels and the exploration of uncertain channels. The simulation results show that compared with the existing attack strategy, the proposed attack strategy can better adapt to the non-stationary cognitive wireless network by optimizing the attack decision through online learning.