Image Source Identifier with Video Deepfake and Forgery Detection using Conventional and Recurrent Neural Networks

Abstract

This paper presents a model which addresses two important forensics analysis i.e. source camera identification and deepfake/forgery detection for videos.Camera identification of images is implemented using CNN based Photo Response Non-Uniformity noise patterns and for Deepfake/Forgery we have used RNN based model that analyzes temporal and spatial inconsistencies. The source identification model is trained on multiple mobile cameras whereas the deepfake model captures the frame-wise variations and deepfake artifacts,The experimental results of the model validate the robustness of the system with an overall approximate of 93% accuracy providing a comprehensive framework for Digital Media Forensics.

Introduction

1. Introduction

With the rise of cloud technologies and media-sharing platforms, verifying the authenticity of digital images and videos has become increasingly difficult. The spread of deepfakes and image forgeries poses significant threats to domains such as cybersecurity, forensics, and journalism.

• Deepfakes use GANs (Generative Adversarial Networks) to generate hyper-realistic but fake videos.

• Forgeries often use tools like Photoshop to alter images maliciously.

• Verifying source authenticity is crucial, especially in legal or investigative contexts.

2. Key Components of the System

???? A. Source Camera Identification

• Every camera sensor has a unique pattern known as PRNU (Photo Response Non-Uniformity) caused by manufacturing imperfections.

• A CNN-based model extracts PRNU noise to attribute images to their source camera.

• Implemented using a 6-layer neural network, trained on the MICHE-I dataset.

• Accuracy: Reached up to 93% depending on model configuration.

???? B. Deepfake Detection

• Implemented using a ResNet-50 CNN architecture on the FaceForensics++ dataset and a custom dataset.

• Video frames are preprocessed, normalized, and then passed through the ResNet model.

• Residual connections in ResNet help detect deepfakes by analyzing frame sequences.

• Accuracy improves with the number of frames used in testing.

3. Methodology

???? Camera Identification Steps

• Extract PRNU noise from each image.

• Feed through a CNN with:

  • 2 Conv2D layers, pooling, flattening, and fully connected layers.

  • Final softmax layer outputs camera classification.

• Categorical cross-entropy is used as the loss function.

???? Deepfake Detection Steps

• Extract video frames.

• Normalize and feed into ResNet-50.

• Output classified as real or fake via softmax layer.

• Accuracy increases with sequence strength (i.e., number of frames used).

4. Technologies Used

• TensorFlow – Model training and evaluation.

• NumPy, Pandas – Data preprocessing and manipulation.

• Matplotlib – Visualization.

• Scikit-learn – Evaluation metrics, classifier implementations.

• OpenCV – Video and frame processing.

5. System Architecture

• Input (images/videos) → Metadata & Frame Extraction → PRNU Noise & Deepfake Analysis → Multimodal Evidence Fusion → Decision Output

• Modular components include:

  • PRNU extraction

  • Deepfake detection using frame-level analysis

  • Integration pipeline for accuracy and consistency

6. Experimental Results

???? Deepfake Detection Accuracy by Frame Count

???? Source Camera Identification Performance

7. Related Work

Summarizes several notable studies in both source camera identification and deepfake detection:

• CNN, RNN, LSTM, XceptionNet, ResNet, and Vision Transformers are commonly used.

• Datasets like DFDC, Celeb-DF, FaceForensics++, QUFVD are standard benchmarks.

•  

Conclusion

In this paper we successfully implement a deep learning application, for the source camera identification and deepfake detection of videos. This addresses the current challenges faced by the digital media authentication sector. Employing a CNN architecture for deepfake and source identification, the model accurately facilitates the information required. The highly efficient deepfake module leveraging ResNet-50 based on CNN architecture with accuracies within the range of 84% - 93% also helps in addressing the concern of subtle artifacts introduced during synthetic content generation. And source camera identification also based on CNN architecture giving the accuracy up to 93% plays a crucial role in forensic investigations, and provenance of digital content. The integration of these modules in a unified analytical pipeline enhances the system’s reliability.

References

[1] “Source Camera Attribution from Videos”- Husrev TahaSencar. Book series of Advances in Computer Vision and Pattern Recognition, 2022 [2] “Deepfake Detection”– Siwei Lyu. Book Series of Advances in Computer and Pattern Recognition, 2022 [3] “Source Camera Device Identification from Videos” – Guru Swaroop Bennabhaktula, Derrick Timmerman, Enrique Alegre, George Azzopardi. SN Computer Science Journal, 2022 [4] “A fast source camera identification and verification method based on PRNU analysis for use in video forensic investigations” – Wen-Chao Yang, JaijunJiang, Chung-Hao Chen. Multimedia Tools and Applications Journal, 2020 [5] “ A Survey on Deepfake Video Detection” - Peipeng Yu, Zhihua Xia, Jianwei Fei, Yujiang Lu. IET Biometrics Journal, 2020 [6] “Recurrent Convolutional Structure for Audio Spoof and Video Deepfake Detection” – Akash Chintha, Bio Thai, Saniat Javid Sohrawardi, Kartavya Bhatt, Andrea Hickerson, Matthew Wright, Raymond Ptucha. IEEE Journal of Selected Topics in Signal Processing, Vol 14, No. 5, 2020 [7] “Deepfake Video Detection Using Transfer Learning Approach” – Shraddha Suratkar, Faruk Kazi. Arabian Journal for Science and Engineering, 2023 [8] “Real-Time Deepfake Video Detection Using Eye Movement Analysis with a Hybrid Deep Learning Approach” –Muhammab Javed, Zhaohui Zhang, Fida Hussain Dahri, Asif Ali Laghari. MDPI Electronics Journal 2024 [9] “Deepfake Video Forgery Detection” – Apurv Jindal. IRE Journals, Volume 6 Issue 12, 2023 [10] “Blind Source Camera Identification” –Mehdi Kharrazi, Husrev T. Sencar, Nasir Memon. IEEE International Conference on Image Processing, 2004 [11] “Enhancing Source Camera Identification Performance With a Camera Reference Phase Sensor Pattern Noise” – Xiangui Kang, Yinxiang Li, Zhenhua Qu, Jiwu Huang. IEEE Transactions on Information Forensics and Security, Vol 7, No. 2, 2012 [12] “Identification of the Source Camera of Images Based on Convolutional Neural Network” – Na Huang, Jingsha He, Xinggang Xuan, Gongzheng Liu, Chengyue Chang. Journal of Digital Investigations, 2018 [13] “Deep learning for Source Camera Identification on mobile devices” – David Freire-Obregon, Fibo Narducci, Silvio Barra, Modesto Castrillon-Santana. Journal of Pattern Recognition Letters, 2018 [14] “Image Based Source Camera Identification using Demosaicking” – Yangjing Long, Yizhen Haung. IEEE International Conference on Image Processing, 2006 [15] “Source Camera Device Identification Based on Raw Images” – Tong Qiao, Florent Retraint, Remi Cogranne, Thanh Hai Thai. IEEE International Conference on Image Processing, 2015 [16] “Source Camera Identification Forensics Based on Wavelet Features” – Bo Wang, Yiping Guo, Xiangwei Kong, Fanjie Meng. Fifth International Conference on Intelligent Information Hiding and Multimedia Signal Processing, 2009 [17] “Combining PRNU and noiseprint for robust and efficient device source identification” – Davide Cozzolino, Francesco Marra, Diego Gragnaniello, Giovanni Poggi, Luisa Verdoliva. EURASIP Journal on Information Security 2020

Copyright

Copyright © 2025 Rahul Daga, Pranav Patil, Parth Rajopadhye, Arihant Awate, Manasvi Patil, Deepali Joshi. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.