Emotion is an internal state of human being that arises due to interpersonal events. It plays an important role in social interactions, learning, perception and decision making. An emotion is a linear combination of affective dimensions namely arousal and valence. The levels of emotional intensity are expressed in terms of arousal dimension. The valence dimension quantifies the pleasantness and unpleasantness of the emotion. Emotion is analyzed using physiological signals such as ECG, EEG and clinical examination. Recording and analysis of Electro Dermal Activity (EDA) signals is a widely-used technique to characterize various emotional states. EDA is a non-invasive technique that records the skin conductivity of emotional sweating. In this work, an attempt has been made to differentiate various emotional states using EDA signals, that are obtained from a publicly available DEAP database. To eliminate the selection of handcrafted features, Deep Belief Network (DBN) is used which automatically extract features from raw EDA signals. It uses unsupervised feature learning architecture to build classifiers and predict the arousal-valence levels for classification. The result shows that DBN classifiers are able to differentiate these different emotional states. This yields an average classification accuracy for arousal and valence dimensions with accuracy 71.25% and 70%, respectively. Thus, it appears that the proposed approach can be used for ambulatory monitoring to differentiate various emotional states. Copyright 2017, ISA All Rights Reserved.