Background: The electroencephalogram (EEG) emotion classification/recognition is one of the popular issues for advanced signal classification. However, it is difficult to manually screen the EEG signals as they are highly nonlinear and non-stationary. Methods: This paper introduces a novel nonlinear and multileveled features-based automatic EEG emotion classification method. Our presented EEG classification model uses feature vector creation deploying an S-Box-based local pattern with a decomposition (tunable q-factor wavelet transform is utilized), the most significant features chosen, classification a shallow machine learning method, and hard majority voting. The novel side of this research is the presented feature extractor since a component of the Clefia cipher has been considered to create a local feature extractor. Results: We have obtained an accuracy of 100.0%, 98.02%, 99.33%, and for valence, arousal, and dominance cases using the DEAP database. Also, we achieved 99.69%, 98.98%, and 99.66% accuracies for valence, dominance, and arousal cases with the DREAMER database. Our proposed model is able to classify arousal, dominance, and valence cases with an accuracy of more than 98% using both databases. Conclusions: The results show that the clefia pattern can perform automatic emotion classification with low computational complexity and high accuracy.