Exposing Manipulated Photos and Videos in Digital Forensics Analysis

Area of Science:

• Digital Forensics
• Cybersecurity
• Machine Learning

Background:

• Tampered multimedia content, including deepfakes, is prevalent in cybercrimes like fake news and ransomware.
• Current criminal investigations face challenges in automatically distinguishing fake from genuine media.
• Automated digital forensics tools are crucial for efficient detection of manipulated content.

Purpose of the Study:

• To develop and implement a machine learning-based method for detecting tampered digital photos and videos.
• To integrate this detection method into the Autopsy digital forensics platform.
• To evaluate the method's effectiveness in identifying deepfake content.

Main Methods:

• A Support Vector Machines (SVM) model was developed to classify multimedia files as genuine or fake.
• Features were extracted using Discrete Fourier Transform (DFT) applied to images and video frames.
• The method was implemented in Python and integrated into Autopsy modules.
• Evaluation involved large datasets of classified photos and videos, including the Celeb-DFv1 dataset for deepfake detection.

Main Results:

• The DFT-SVM method achieved high average F1-scores: 99.53% for photos, 79.55% for videos, and 89.10% for mixed content.
• Performance surpassed existing SVM-based approaches in literature.
• While Convolutional Neural Networks (CNNs) showed higher accuracy, the DFT-SVM method demonstrated competitive results with significantly reduced processing time.

Conclusions:

• The proposed DFT-SVM method is a viable and efficient tool for detecting tampered multimedia content within digital forensics.
• Its integration into Autopsy facilitates automated detection and aids criminal investigators.
• The method provides a practical solution for identifying deepfakes, balancing accuracy with computational efficiency.