DeFake AI: A Machine Learning based Approach to Audio-Video Deepfake Detection

The fast expansion in deepfake technology has raised the concerns regarding digital security, the reliability of information, and the overall stability of people’s trust. This paper illustrates DeFake AI, a detection system developed to examine fake/generated images, videos, and audios through a mix of deep learning methods. Its framework uses a MobileNet–LSTM arrangement for visual assessment and a CNN-based setup for audio, supported with TensorFlow, OpenCV, Librosa and Streamlit. The system processes mel-spectrogram inputs for audio signals and identifies temporal–spatial changes within visual media, along with motionguided frame selection and a confidence-weighted merging of predictions. The web interface is designed to provide real-time, multimodal checks. Tests indicate consistently strong detection accuracy, and the addition of confidence scoring adds steadiness to the classification process; overall, DeFake AI acts as a practical and flexible solution for verifying media authenticity while addressing broader issues in digital forensics and misinformation control.