For natural language processing tasks like sarcasm detection, the growing amount of user-generated content on social media platforms offers significant challenges. Sarcastic content frequently uses noisy expressions and nuanced language, which reduces the effectiveness of traditional classifiers. To improve sarcasm detection in noisy textual environments, this paper suggests a hybrid structure for the combination of deep neural network with wrapper-based feature selection and fuzzy logic-based noise alleviation. We first used fuzzy logic to remove low-confidence samples based on probabilistic thresholds using a benchmark dataset of sarcastic headlines. Both classical and hybrid machine learning models then used the feature space after it was refined by wrapper-based recursive feature elimination (RFECV). In a cleaned dataset, our analysis through comparison shows that the proposed deep neural network model with fuzzy and wrapper techniques performed considerably better than more conventional models like SVM (95.71%), logistic regression (95.62%) and naive Bayes (94.57%), with an accuracy of 95.74%. The outcomes validate how well fuzzy and wrapper logic work to minimize noise and maximize feature representation in sarcasm detection tasks.