Forensic Analysis for Detecting Deep-Fake Images Using A Convolutional Neural Network (CNN) and The National Institute of Standards and Technology (NIST) Methods

Abstract

The development of Artificial Intelligence (AI) has significantly influenced audio, video, and image manipulation techniques, commonly known as deepfakes. Image forensics faces an urgent challenge in identifying and mitigating the impact of deepfake content to maintain the integrity and credibility of digital information. This research aims to perform forensic analysis in accordance with NIST standards and to implement Convolutional Neural Network (CNN) methods to detect deepfake images. Based on the test results, the Convolutional Neural Network (CNN) method can be effectively applied to deepfake image detection. The CNN architecture used can identify the distinct visual characteristics of deepfake images with high performance. The model demonstrates the ability to learn and minimize prediction errors on training data. Accuracy graphs indicate that the model has successfully learned data patterns, as evidenced by consistent improvements in both training and validation data as the number of epochs increases. Furthermore, the loss graph shows a downward trend, signifying a continuous reduction in model error. The precision graph demonstrates the model's effectiveness in reducing false positives, thereby minimizing errors in detecting the original data. The recall graph also indicates improved detection performance on the training data. The ROC curve suggests that the model possesses superior classification capabilities compared to random guessing. Additionally, the Area Under the Curve (AUC) of 0.6544 serves as a quantitative indicator of performance, indicating that the model has moderate capability for class differentiation. Detection results from the CNN model on a dataset of real and deepfake images show that the Confidence and Raw Score values can distinguish between the two; however, the confidence levels still fluctuate around the classification threshold. Low confidence values in certain images suggest that the extracted features are not yet optimal at distinguishing between real faces and manipulated images. Moreover, the application of the National Institute of Standards and Technology (NIST) standards (Collection, Examination, Analysis, and Reporting) for forensic analysis ensures that the evidence gathered is legally accountable in court. Thus, these standards can serve as a scientific reference to ensure a more structured and standardized investigation process for deepfake images.