Abstract: ‍ ‍Modulation classification techniques based deep learning are gradually being integrated and applied with their advantages in feature extraction and recognition performance. However， due to the vulnerability and limitation of the intelligent spectrum sensing model， the performance can be reduced by adding imperceptible perturbation to the original signal. Aiming at the vulnerability of intelligent electromagnetic signal modulation classification system， this paper carried out the research of electromagnetic signal deception method. Different from adding global disturbance on the signal， this method added disturbance to the specific sampling points， and the electromagnetic signal modulation classification system can be deceived smartly. The results show that this method could produce a relatively slight invisible disturbance， which deceives the system produce machine hallucinations and identify adversarial example as targeted signal and disturb the identification result of the system， so as to achieve the goal of “concealing the true and revealing the false” .