Emotions are fundamental for human decision-making and interactions which influences all the different aspects of our day-to-day life. Psychologists have developed different types of models for understanding the different complex emotional steps. With technological advancement, automated emotion recognition has become increasingly significant, with applications spanning education, marketing, human-robot interaction, and healthcare. Emotion recognition can be approached through two different verticals: uni-modal and multimodal methods. Each vertical has its own unique benefits. Physiological signals derived from various modalities provide important information relating human emotions. As the physiological signals are generated by the central nervous system, so human beings do not have any control over it. This research work mainly focuses on classification of emotional steps using a non-invasive physiological signal, electrocardiogram (ECG) signals from a standard dataset. ECG records the electrical activity of the heart over a period of time and has a deep relation with the emotional state of the person. In this work, emotions are categorized on the valence-arousal plane into four states using standard supervised machine learning algorithms. The work has been done in two phases: The first approach was a class imbalanced one, to overcome the challenges of class imbalance, in the second approach Generative Adversarial Network (GAN) was used to generate the synthetic samples to balance the class distributions. This novel approach allowed for effective application of utilizing conventional supervised machine learning techniques on physiological signals for identification of emotion in four quadrants of valence-arousal plane. Out of the different techniques, KNN outperformed other supervised approaches.