Focusing on the electromagnetic information security of keyboard, the working principle and signal characteristics of PS/2 keyboard are analyzed, and a detection method based on deep learning is put forward. This method adapts the Convolution Neural Network(CNN) structure to the electromagnetic leakage signal of the keyboard equipment. Combining with the improved Gradient weighted Class Activation Mapping(Grad-CAM) method, the intelligent recognition and location of keyboard electromagnetic information characteristics are realized. Testing on the electromagnetic signals of four keyboards, the classification accuracy can reach 98%. And the classification accuracy can reach 81% in noisy environment. The experimental results indicate that the improved method has better location performance for the electromagnetic information of keyboards.