Enhancing Digital Evidence Analysis: Fakeseeker for Real-Time Deepfake Detector using EfficientNet

Abstract

Deepfake technology poses significant challenges to the authenticity of digital evidence. This creates an urgent need for reliable and effective verification methods. This work introduces Fakeseeker, a lightweight, real-time detection framework designed to assist digital forensic investigations. The system uses a fine-tuned EfficientNet-B0 architecture to analyze both uploaded and live-streamed media. The approach involves preparing datasets from publicly available sources, including Celeb-DF, FaceForensics++, and DFDC. Subsequently, MTCNN-based face extraction, data augmentation, and targeted optimization of EfficientNet-B0 were undertaken. Comparative evaluations with EfficientNet variants B0, B1, and B2 show that EfficientNet-B0 offers the best balance between accuracy and efficiency. It takes about nine minutes per training epoch, achieves a validation LogLoss of 0.0021, and has an AUC score of 0.9996. Based on these results, Fakeseeker was built using EfficientNet-B0 as its backbone. It includes separate components for media processing, classification, and reporting. Evaluation results show a detection accuracy of 99.2% with minimal resource usage. This makes the system suitable for real-time forensic workflows and field applications. Overall, this study provides a practical and scalable solution for improving the verification and analysis of digital evidence.