Deepfake Forensic Analysis: Source Dataset Attribution and Legal Implications of Synthetic Media Manipulation

View on arXiv ← Back to list

Authors: Massimiliano Cassia, Luca Guarnera, Mirko Casu, Ignazio Zangara, Sebastiano Battiato

Published: 2025-05-16 10:47:18+00:00

AI Summary

This paper proposes a forensic framework for identifying the training dataset of GAN-generated images using interpretable features like spectral transforms, color distribution metrics, and local feature descriptors. Supervised classifiers achieve high accuracy (98-99%) in both binary (real vs. synthetic) and multi-class dataset attribution.

Abstract

Synthetic media generated by Generative Adversarial Networks (GANs) pose significant challenges in verifying authenticity and tracing dataset origins, raising critical concerns in copyright enforcement, privacy protection, and legal compliance. This paper introduces a novel forensic framework for identifying the training dataset (e.g., CelebA or FFHQ) of GAN-generated images through interpretable feature analysis. By integrating spectral transforms (Fourier/DCT), color distribution metrics, and local feature descriptors (SIFT), our pipeline extracts discriminative statistical signatures embedded in synthetic outputs. Supervised classifiers (Random Forest, SVM, XGBoost) achieve 98-99% accuracy in binary classification (real vs. synthetic) and multi-class dataset attribution across diverse GAN architectures (StyleGAN, AttGAN, GDWCT, StarGAN, and StyleGAN2). Experimental results highlight the dominance of frequency-domain features (DCT/FFT) in capturing dataset-specific artifacts, such as upsampling patterns and spectral irregularities, while color histograms reveal implicit regularization strategies in GAN training. We further examine legal and ethical implications, showing how dataset attribution can address copyright infringement, unauthorized use of personal data, and regulatory compliance under frameworks like GDPR and California's AB 602. Our framework advances accountability and governance in generative modeling, with applications in digital forensics, content moderation, and intellectual property litigation.

Key findings

The classifiers achieved 98-99% accuracy in distinguishing real from synthetic images and attributing synthetic images to their source datasets. Frequency-domain features (DCT/FFT) proved highly effective in capturing dataset-specific artifacts.

Approach

The framework extracts discriminative statistical signatures from synthetic images using spectral transforms (Fourier/DCT), color distribution metrics, and local feature descriptors (SIFT). These features are then used to train supervised classifiers (Random Forest, SVM, XGBoost) for dataset attribution.

Datasets

CelebA, FFHQ; GAN-generated images from StyleGAN, AttGAN, GDWCT, StarGAN, and StyleGAN2 trained on CelebA or FFHQ.

Model(s)

Random Forest, SVM, XGBoost

Author countries

Italy

← Previous