arxiv_cv 95% Match Research Paper Cybersecurity Professionals,Digital Forensics Experts,AI Researchers,Content Moderation Teams 2 days ago

Who Made This? Fake Detection and Source Attribution with Diffusion Features

generative-ai › diffusion

📄 Abstract

Abstract: The rapid progress of generative diffusion models has enabled the creation of synthetic images that are increasingly difficult to distinguish from real ones, raising concerns about authenticity, copyright, and misinformation. Existing supervised detectors often struggle to generalize across unseen generators, requiring extensive labeled data and frequent retraining. We introduce FRIDA (Fake-image Recognition and source Identification via Diffusion-features Analysis), a lightweight framework that leverages internal activations from a pre-trained diffusion model for deepfake detection and source generator attribution. A k-nearest-neighbor classifier applied to diffusion features achieves state-of-the-art cross-generator performance without fine-tuning, while a compact neural model enables accurate source attribution. These results show that diffusion representations inherently encode generator-specific patterns, providing a simple and interpretable foundation for synthetic image forensics.

Authors (3)

Simone Bonechi

Paolo Andreini

Barbara Toniella Corradini

Submitted

October 31, 2025

arXiv Category

cs.CV

arXiv PDF

Key Contributions

FRIDA is a lightweight framework for deepfake detection and source attribution that leverages internal activations from pre-trained diffusion models. It achieves state-of-the-art cross-generator performance without fine-tuning using a k-NN classifier on diffusion features, and enables accurate source attribution with a compact neural model, demonstrating that diffusion representations encode generator-specific patterns.

Business Value

Enhances trust in digital media by providing tools to identify synthetic content and its origin, crucial for combating misinformation and protecting intellectual property.

Paper Metadata

Innovation Type

Methodological Innovation

Deployment Feasibility

High, as it leverages pre-trained models and lightweight classifiers, making it efficient for real-time analysis.

Limitations Addressed

Poor generalization of supervised detectors to unseen generators,Need for extensive labeled data and frequent retraining,Difficulty in attributing synthetic images to their source generator

Performance Gains

State-of-the-art cross-generator performance

Technical Tags

deepfake detectionsource attributiondiffusion modelsgenerative modelsinternal activationscross-generator generalizationk-nearest-neighborneural network

Research Topics

Generative AI SecurityComputer VisionDeepfake DetectionSource AttributionMachine Learning

Methods & Architectures

FRIDA (Fake-image Recognition and source Identification via Diffusion-features Analysis)Diffusion Featuresk-Nearest-Neighbor (k-NN) classifierCompact Neural Model Pre-trained Diffusion Modelsk-NN ClassifierNeural Network

Applications & Tasks

Digital Forensics Content Authenticity Cybersecurity Generalization to Unseen GeneratorsDeepfake DetectionSource Generator AttributionNeed for Retraining Deepfake DetectionSynthetic Image Source Attribution

Related Fields

Computer VisionGenerative Adversarial Networks (GANs)Digital ForensicsCybersecurity

Keywords

deepfake detectionsource attributiondiffusion modelsgenerative AIsynthetic imagesimage forensicsk-NNneural networkscontent authenticitymisinformation

Academic Context

#Generative AI Security#Computer Vision#Deepfake Detection#Source Attribution#Machine Learning

Commercial Potential

Potential Products

Deepfake detection toolsContent provenance verification servicesAI-generated content analysis platforms

Target Industries

Media & EntertainmentSocial MediaCybersecurityGovernment

Use Case Examples

Verifying the authenticity of news imagesDetecting manipulated media in social networksTracing the origin of AI-generated propaganda

Competitive Edge

Offers a novel approach using diffusion model internals, achieving better cross-generator generalization than existing methods without requiring fine-tuning on new generators.

Market Opportunity

Large and growing, due to the increasing prevalence of synthetic media.

Revenue Models

SaaS for content verificationlicensing to media platforms.

Resource Requirements

Compute Needs

Low to moderate, leverages pre-trained models and lightweight classifiers.

Data Requirements

Diverse datasets of real and synthetic images from various generators.

Deployment Constraints

Requires access to pre-trained diffusion models (or their internal representations).

Scalability

Scalable due to the lightweight nature of the k-NN and compact neural model.

Regulatory Considerations

Potential for misuse in surveillance or censorship

Production Readiness

Maturity Level

Research

Time to Market

1-2 years for productization.

Patent Potential

Moderate, for the FRIDA framework and its application.

View Full Paper Back to Papers