The rapid proliferation of deepfake videos—synthesised using Generative Adversarial Networks (GANs) and allied\ndeep-learning techniques—poses grave risks to societal trust, democratic processes, and personal privacy. Existing detection\napproaches predominantly rely on frame-level spatial analysis and consequently fail to capture temporal inconsistencies that\narise in manipulated sequences. This paper presents Deep Shield, a hybrid deep-learning framework that couples a ResNeXt\nconvolutional neural network (CNN) for spatial feature extraction with a Long Short-Term Memory (LSTM) recurrent network\nfor temporal sequence modelling. Each video frame is first preprocessed via face detection and alignment, after which ResNeXt\nencodes per-frame spatial embeddings that are subsequently fed into the LSTM to capture inter-frame inconsistencies. A fully\nconnected classifier then labels the video as Real or Fake alongside a confidence score. The system is validated on three\nbenchmark datasets—FaceForensics++, DFDC, and Celeb-DF—achieving detection accuracy exceeding 99 % together with\nprecision, recall, and F1-score values above 99 %. The framework is wrapped in a Django-based web interface that allows nontechnical users to upload videos and obtain results in near real time. Robustness testing under compression artefacts, low-light\nconditions, and adversarial inputs confirms the generalisability of the approach.