Sarcasm is a prevalent type of symbolic language that is often hard to identify with conventional natural language processing methods when it comes to social media platforms. There may be difficulty identifying it, particularly in brief writing such as messages on social media without context. Sarcasm detection technologies may not always be accurate since their accuracy depends on the type and volume of data used to train the models. In order to overcome this problem, the authors propose an attention-based RoBERTa model. In this paper, a transfer learning model, RoBERTa with attention, has been proposed to detect sarcasm in social media. DL models using the iSarcasm dataset are evaluated using metrics of accuracy and F1-score. With an F1 score of 75.99%, our model outperforms in identifying sarcasm in social media and can help create more precise and trustworthy NLP tools for social media data analysis.