AI-Based Deepfake Detection with Sentiment Analysis

The work "AI-Based Deepfake Detection with Sentiment Analysis" addresses the growing concern of digital media manipulation. Deepfakes, i.e. synthetic or manipulated videos, threaten the authenticity of online content, fuelling misinformation, social manipulation and an erosion of trust in digital platforms. The proposed web-based system combines two functions: deepfake detection with a CNN–LSTM network (a ResNet-18 frame encoder followed by an LSTM that models temporal inconsistencies across frames), and sentiment analysis with a hybrid text–visual method that fuses lexicon-based scoring of the video description with visual cues (brightness and colour warmth), while extracting BERT and ResNet-18 features for future learned fusion. Uploaded videos are validated, frames are extracted with OpenCV, and the results are presented in an interactive report with confidence scores and sentiment insights that can be exported as a PDF. During model selection, candidate deepfake backbones were benchmarked on FaceForensics++, where the best candidate (InceptionResNetV2) reached 94% accuracy. In a preliminary evaluation of the deployed system on 15 videos from three content categories, both the deepfake detector and the sentiment module classified 12 of the 15 videos correctly (80%), and a usability survey with 10 participants rated upload ease and interface design favourably while identifying processing speed as the main limitation. By reporting both the authenticity and the emotional tone of a video in one accessible tool, this work aims to support media literacy and responsible content verification. Source code: https://github.com/AsjadS18/AI-Based-Deepfade-detection-with-Sentiment-analysis