Abstract: To authenticate cigarette effectively and conveniently, an authentication method based on the morphology of the glue lines on cigarette packets was proposed. An image acquisition device was adopted to capture glue line images on the packet labels. The background that had nothing to do with the detection in the images was removed to obtain the glue line image samples. Based on the deep learning anomaly detection method, only normal samples were used to train the convolutional neural network classification model, and the maximum prediction probability generated from the samples to be tested with the model was used for threshold discrimination. Five classical convolutional neural network models were tested separately on authentic cigarette packets produced by 10 types of cigarette packers and counterfeit cigarette packets involving about 60 cigarette specifications, and a total of 1 294 cigarette packet samples were identified. The results showed that: 1) All models were capable of identifying authenticity to some extents when the average of the maximum prediction probability values of the test set samples was used as the threshold. 2) The identification accuracy of the ResNet18, DenseNet121, and InceptionV1 models was approximately 94%, the recall rate was approximately 93% to 94%, the specificity was approximately 93% to 95%, and the AUC (Area Under Curve) was approximately 0.98. InceptionV1 model featured less parameters and easier deployment. 3) This method could construct an authentication model without counterfeit samples and was suitable for cigarettes of various specifications. This technique provides support for the establishment of efficient and reliable methods for the intelligent identification of cigarette authenticity.