The global availability of false information and fake news across digital platforms endangers public opinion, democratic integrity, and the stability of society. In recognition of this real world challenge, this thesis proposes a hybrid deep learning based model designed for detecting fake news that takes into account both textural and social contextual features such as engagement metrics (likes, shares) and source credibility. As suggested, the system is composed of a parallel neural network that uses user-level social metadata to incorporate engagement metrics and An LSTM network that acquires knowledge from the language patterns found in news articles. This dual-input architecture enables the model to capture both semantic cues and social behavior characteristics associated with misinformation. The WELFake dataset is used for training and evaluation, and the performs well, surpassing conventional machine learning benchmarks. The research also includes a comparative analysis of various model configurations, highlighting the effectiveness of multi-source information fusion in enhancing detection performance. The proposed approach provides a scalable, real-time, and interpretable solution for significantly reduce the impact of fake news on digital platforms.