Face-trace: Open-Set Attribution and Progressive Discovery of Synthetic Face Generators

2026-07-08 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy · Depth: Expert, quick

Summary

Face-trace is a novel pipeline designed for open-set synthetic face source attribution, addressing the challenge of identifying generators for images produced by both known and previously unseen models. Unlike traditional closed-set approaches, Face-trace integrates known generator classification, energy-based Out-of-Distribution (OOD) rejection, and unknown generator discovery. The system trains a classifier on known generators using frozen I-JEPA embeddings. For rejected samples, it combines projected I-JEPA features with Forensic Self-Descriptors, then clusters these to identify new, unknown generator groups. This method also extends to an incremental scenario, progressively discovering new generators as samples arrive over time. Experiments on the WILD dataset demonstrate 96.73% closed-set attribution accuracy. In open-set scenarios, energy-based rejection achieved 71.25% balanced accuracy, with clustering yielding an ARI of 0.81, NMI of 0.90, and purity of 87.74%. The incremental setting maintained a final purity of 99.23%.

Key takeaway

For multimedia forensics analysts tasked with identifying the origins of synthetic face images, Face-trace offers a robust approach to handle both known and novel generative models. You should consider integrating open-set attribution pipelines like Face-trace to move beyond closed-set limitations, enabling the discovery and organization of previously unseen synthetic content generators. This capability is crucial for tracking the evolving landscape of deepfake creation and improving content provenance analysis.

Key insights

Face-trace enables open-set attribution of synthetic faces by combining known generator classification with energy-based OOD rejection and unknown generator discovery.

Principles

• Open-set attribution requires OOD rejection.
• Combine feature embeddings for robust representation.
• Incremental discovery extends generator space.

Method

Train a classifier on known generators using frozen I-JEPA embeddings. Represent rejected samples with combined I-JEPA features and Forensic Self-Descriptors, then cluster them to discover unknown generator groups incrementally.

In practice

• Identify sources of deepfake images.
• Track emerging synthetic content generators.
• Organize previously unseen synthetic samples.

Topics

• Synthetic Face Attribution
• Open-Set Recognition
• Generative AI Forensics
• I-JEPA Embeddings
• Out-of-Distribution Detection
• Deepfake Detection

Best for: Research Scientist, AI Scientist, Computer Vision Engineer, AI Security Engineer

Related on AIssential

• ForensicConcept: Transferable Forensic Concepts for AIGI Detection — ForensicConcept enables transferable forensic concept extraction for robust AI-generated image detection across unseen generators.
• Can liveness detection models generalise to synthetic media generation techniques they were never trained on? [D] — Liveness detection models fail to generalize to new deepfake techniques due to learning outdated generation artifacts.
• Anil-matcha / Open-Generative-AI — Open Generative AI offers an uncensored, open-source platform for AI media generation with extensive model support and local inference capabilities.
• SEAL: Semantic-aware Single-image Sticker Personalization with a Large-scale Sticker-tag Dataset — SEAL improves single-image sticker personalization in diffusion models by mitigating overfitting and enhancing contextual control.
• Multi-Frequency Fusion for Robust Video Face Forgery Detection — Lightweight fusion of handcrafted cues can significantly improve face forgery detection accuracy with minimal model overhead.
• Proto-LeakNet: Towards Signal-Leak Aware Attribution in Synthetic Human Face Imagery — Proto-LeakNet exploits diffusion model "signal leaks" in latent space for robust, interpretable synthetic image source attribution.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.