Cross-eye jamming is an electronic attack technique that induces an angular error in the monopulse radar by artificially creating a false target and deceiving the radar into detecting and tracking it. Presently, there is no effective anti-jamming method to counteract cross-eye jamming. In our study, through detailed analysis of the jamming mechanism, a multi-target model for a cross-eye jamming scenario is established within a random finite set framework. A novel anti-jamming method based on multitarget tracking using probability hypothesis density filters is subsequently developed by combining the characteristic differences between target and jamming with the releasing process of jamming. The characteristic differences between target and jamming and the releasing process of jamming are used to optimize particle partitioning. Particle identity labels that represent the properties of target and jamming are introduced into the detection and tracking processes. The release of cross-eye jamming is detected by estimating the number of targets in the beam, and the distinction between true targets and false jamming is realized through correlation and transmission between labels and estimated states. Thus, accurate tracking of the true targets is achieved under severe jamming conditions. Simulation results showed that the proposed method achieves a minimum delay in detection of cross-eye jamming and an accurate estimation of the target state.