A Real-Time Edge-AI Hardware Device for Multimodal Deepfake Detection and Authenticity Verification
Abhirami J S
Assistant Professor
Department of Artificial Intelligence and Data Science Nehru Institute of Engineering and TechnologyCoimbatore, Tamil Nadu, India
nietjsabhirami@nehrucolleges.com
Rishikesh K
Department of Artificial Intelligence and Data Science
Nehru Institute of Engineering and Technology
Coimbatore, Tamil Nadu, India
workwithrishikeshk@gmail.com
Celestian A
Department of Artificial Intelligence and Data Science
Nehru Institute of Engineering and Technology
Coimbatore, Tamil Nadu, India
celestianarojack@gmail.com
Sujitha S
Department of Artificial Intelligence and Data Science
Nehru Institute of Engineering and Technology
Coimbatore, Tamil Nadu, India
sujithasuresh803@gmail.com
Abstract— The rapid advancement of generative artificial intelligence has enabled the creation of highly realistic synthetic media referred to as deepfakes. These manipulated videos and voices can imitate real individuals and are increasingly leveraged for misinformation, identity fraud, and cybercrime. Existing deepfake detection systems are primarily software-based and cloud-dependent, limiting their accessibility, privacy, and real-time usability. This paper proposes TrustLens, a portable Edge-AI hardware device designed to detect manipulated multimedia content through multimodal analysis of both audio and video signals. The system integrates a Raspberry Pi 5-based embedded computing platform with real-time media capture hardware and deep learning models capable of identifying facial manipulation artifacts and voice cloning characteristics simultaneously. The device performs on-device inference ensuring low latency and privacy preservation without cloud reliance. TrustLens analyzes visual artifacts including facial texture inconsistencies, abnormal blinking patterns, and lip synchronization mismatches, while evaluating spectral anomalies and unnatural prosody associated with synthetic speech. Detection outputs are fused using weighted aggregation to produce a unified authenticity trust score. Experimental evaluation demonstrates that multimodal analysis achieves 93.6% accuracy, significantly outperforming single-modality baselines. The system offers practical applications in banking security, digital forensics, journalism verification, and cybersecurity monitoring.
Keywords—Deepfake Detection, Edge Artificial Intelligence, Multimodal Analysis, Embedded Systems, Voice Cloning Detection, Cybersecurity, Real-Time Inference.
