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

TL;DR

This work tackles the provenance problem of GAN-generated media by proposing a multimodal forensic framework that attributes synthetic images to their training datasets. It combines frequency-domain ($DCT$, $FFT$), color distribution, and local texture features ($SIFT$) with supervised classifiers to perform real-vs-synthetic discrimination and dataset attribution across multiple GAN architectures. The approach yields high accuracy and reveals that frequency-domain artifacts carry strong dataset-specific fingerprints, enabling governance actions for copyright, privacy, and regulatory compliance under frameworks like GDPR and AB 602. By linking synthetic outputs to their training data, the framework advances accountability in generative modeling with practical implications for digital forensics, content moderation, and intellectual-property litigation.

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.

Figures (4)

• Figure 1: Methodological Pipeline for Dataset Attribution Using Multimodal Features and Machine Learning Classifiers. The pipeline illustrates the process from dataset selection (including GAN-generated images from CelebA and FFHQ) to feature extraction (using DCT, RGB histograms, FFT, and SIFT) and classification using Random Forest, SVM, K-NN, and XGBoost. The final step involves attributing synthetic images to their respective source datasets (CelebA or FFHQ).
• Figure 2: Average DCT coefficient heatmaps for real and GAN-generated images. Notable energy dispersion and spatial artifacts distinguish real (CelebA, FFHQ) from synthetic distributions (GDWCT, AttGAN, StarGAN, StyleGAN, StyleGAN2).
• Figure 3: Mean FFT power spectra for real and GAN-generated images. Real datasets (CelebA, FFHQ) exhibit symmetrical low-frequency concentration, while synthetic distributions (GDWCT, AttGAN, StarGAN, StyleGAN, StyleGAN2) display cross-shaped high-frequency components.
• Figure 4: Average RGB color histograms. Real datasets maintain characteristic channel skews, while GAN-generated images display color balancing patterns.