Source-linked AI summary
Source-Face Authenticity Detection for 3D Gaussian Heads Reconstructed from a Single Portrait: A Benchmark and Dedicated Detector
Yujie Gao, Zijian Yu, Yan Hong, Jun Lan, Jianfu Zhang
TL;DR
The paper addresses authenticity detection for 3D Gaussian heads, where reconstruction can weaken forgery cues and alter their visibility across views. It introduces a large-scale benchmark and a two-stage detector combining detail preservation, cross-view consistency, and multi-level classification, achieving the best results across reported metrics.
Problem
3D Gaussian head reconstruction can compress or alter subtle, localized forgery cues and change their visibility across views, while existing detectors lack mechanisms to preserve fine-grained information and cross-view consistency.
Method
The detector uses masked autoencoding and multi-view contrastive learning in Stage I, then freezes the backbone and fuses low-, middle-, and high-level CLS tokens for classification.
Results
Our method consistently outperforms existing detectors, achieves the best result across all evaluation metrics, and ranks first across every reported source category and reconstruction pipeline.
Takeaways & Limitations
The benchmark and detector provide a dedicated framework for evaluating and detecting real versus fake 3D Gaussian heads across multiple sources, views, and reconstruction pipelines.
Takeaways & Limitations
The reconstruction objective does not directly supervise authenticity labels or artifact locations, so its information-preserving benefit remains a hypothesis evaluated through component ablation.
Abstract
from arXiv · showhide
Recent advances in single-image 3D Gaussian head reconstruction have enabled highly realistic and freely renderable digital heads from a single portrait. However, reconstruction and rendering can weaken the forgery traces in the source portrait, making the resulting 3D face difficult to classify whether its underlying face is real or fake, and thereby posing risks to identity authentication and face privacy. To study this problem, we introduce the first large-scale benchmark for this task by collecting real portraits and fake portraits from multiple sources and evaluate representative existing detectors on this benchmark, revealing their lack of explicit mechanisms for retaining fine-grained information and maintaining feature consistency across rendered views. To directly address these two limitations, we propose a detector trained with a two-stage strategy. In Stage I, masked autoencoding encourages the visual backbone to retain the fine-grained appearance information required for local reconstruction, while multi-view contrastive learning enforces feature consistency across rendered views of the same head. Since CLS tokens at different depths exhibit complementary spatial attention patterns, Stage II freezes the adapted backbone and concatenates low-, middle-, and high-level CLS tokens for classification. Experiments show that our method achieves the highest accuracy and ranks first across all reported metrics among the evaluated detectors.
1 Introduction
Single-image 3D Gaussian heads create a new authenticity-detection problem because reconstruction can weaken localized forgery cues and viewpoint changes alter their appearance. The paper introduces a benchmark and a two-stage detector designed to preserve fine-grained information and enforce cross-view consistency.
- Motivation: 3D Gaussian heads reconstructed from portraits can become persistent, freely renderable identity assets that threaten authentication and face privacy.Existing detectors mainly target 2D images or videos, leaving this 3D setting without a dedicated method.
- Challenges: Reconstruction and re-rendering weaken forgery traces into subtle local evidence, while viewpoint changes alter their visibility, scale, and projection.These effects produce different local evidence patterns across views of the same head.
- Benchmark: The benchmark combines forged faces from 9 methods, genuine portraits from 5 datasets, 5 reconstruction methods, and approximately 360K multi-view samples.It evaluates representative face-forgery and general fake-image detectors under a unified benchmark.
- Method: The proposed detector uses masked autoencoding and multi-view contrastive learning before concatenating low-, middle-, and high-level CLS tokens for classification.The first stage targets detail preservation and cross-view consistency; the second stage performs single-view real/fake prediction.
2 Related Works
Prior work spans 3D head representations and 2D authenticity detection, but this paper frames 3D Gaussian head detection as requiring a dedicated two-stage approach. Its framework combines self-supervised representation learning with multi-level feature fusion.
- 3D Gaussian Head Reconstruction: 3D Gaussian Splatting is presented as a recent explicit and efficiently renderable representation for reconstructing heads from a single image.Earlier approaches used parametric face models, implicit representations, tri-plane features, or NeRF-based radiance fields.
- Proposed Framework: The framework is organized into self-supervised fine-tuning followed by multi-level feature fusion and classification.This two-stage structure is the paper’s central design for adapting the backbone and making authenticity predictions.
- Proposed Framework: Stage 1 combines masked autoencoding for local texture details and Gaussian artifacts with multi-view contrastive learning to suppress view-specific shortcuts.Stage 2 normalizes and concatenates low-, middle-, and high-level CLS tokens before MLP classification.
3 Methodology
The methodology constructs a multi-source 3D Gaussian head authenticity benchmark and trains a two-stage detector to preserve fine-grained evidence and consistency across rendered views.
- Dataset Construction: The benchmark reconstructs forged or genuine portraits into renderable 3D identity assets using multiple single-image 3D head methods.The construction includes real portraits, forged portraits, and five reconstruction pipelines.
- Stage I: Self-Supervised Training: Stage I combines masked autoencoding with multi-view contrastive learning to preserve local appearance information and promote view-consistent representations.Masked autoencoding reconstructs masked patches from visible patches, while contrastive learning groups views from the same head and separates different identity groups.
- Stage I: Self-Supervised Training: The contrastive branch treats other views of the same identity group as positives and all non-anchor views as candidates for comparison.This design targets viewpoint-invariant authenticity features across rendered views.
- Stage II: Feature Fusion and Classification: Stage II fuses normalized low-, middle-, and high-level CLS tokens for real/fake classification.The multi-depth representations are concatenated before an MLP produces the prediction and cross-entropy optimization is applied.
- Benchmark Evaluation: The benchmark evaluates authenticity detectors across 2D forgery categories and 3D reconstruction methods.The supplied evaluation materials include comparisons across source categories and reconstruction methods, plus OOD evaluation and baseline comparisons.
4 Experiment
Experiments evaluate the detector on an identity-disjoint benchmark, ablations, representation analyses, and held-out forgery or reconstruction methods. The method achieves the strongest reported performance while masked reconstruction, contrastive learning, and multi-level CLS fusion each support its design.
- Experiment Settings: 147,900 images form an identity-disjoint benchmark with 117,900 training, 15,000 validation, and 15,000 test images.Real and fake samples follow an exact 1:1 ratio within each split.
- Comparison with Baselines: The benchmark evaluates four general-purpose image detectors and three face-forgery or deepfake detectors under a unified protocol.
- Comparison with Baselines: Our method achieves the best result across all reported metrics, ranks first across source categories and reconstruction pipelines, and generalizes best to four held-out subsets.The held-out evaluation excludes the newest method in each forgery category and an unseen 3D reconstruction method from training.
- Ablative Analyses: Removing masked autoencoding or multi-view contrastive learning consistently degrades evaluation metrics, supporting both Stage I objectives.Masked reconstruction also produces plausible facial structure and local details with 75% of input patches masked.
- Ablative Analyses: Contrastive learning mainly improves intra-head similarity and produces more compact same-head clusters with clearer inter-group separation.These analyses support stronger consistency across rendered views of the same head.
- Ablative Analyses: CLS tokens from different depths contain useful authenticity information, while low–middle–high fusion outperforms single-layer averages and the last-layer classifier.Distinct attention regions across depths provide qualitative evidence that their spatial responses are not identical.
5 Conclusion
The paper introduces a large-scale benchmark and a two-stage detector for real/fake 3D Gaussian head detection. The method consistently outperforms existing detectors, but expression-driven forgeries remain challenging.
- The benchmark and detector target real/fake 3D Gaussian head detection using detail-preserving, cross-view-consistent representations and multi-level CLS-token fusion.
- The method consistently outperforms existing detectors across the reported evaluation.
- Expression-driven forgeries remain challenging, with an accuracy of 76.24%.