ARCHIVES

The increasing realism of synthetic media generated using deep learning has intensified the threat posed by deepfake videos in domains such as social media, journalism, legal evidence, and digital identity verification. Existing deepfake detection systems often focus on a single modality, thereby limiting their robustness against sophisticated multimodal manipulations. This paper presents Deep Guard AV, an audio-visual deepfake detection framework that jointly analyzes manipulated video and extracted speech signals while preserving textual transcripts for forensic logging and interpretability. The proposed framework processes video inputs through a dual-stream pipeline. Visual frames are analyzed using a Convolutional Neural Network–Long Short-Term Memory (CNN-LSTM) architecture to capture spatial and temporal inconsistencies, while the extracted audio is processed using a hybrid deep learning model combining waveform-based and spectrogram-based representations. In parallel, the speech content is transcribed and saved locally to maintain an auditable forensic record of processed media. A weighted fusion strategy combines the outputs of audio and video models to produce the final authenticity score. Experimental evaluation demonstrates that integrating audio and video modalities improves detection robustness compared to unimodal analysis. The proposed framework provides an effective and scalable solution for practical deepfake forensics while enhancing transparency through transcript preservation.