Bi-Level Optimization for Self-Supervised AI-Generated Face Detection

Conference: ICCV 2025 arXiv: 2507.22824 Code: github.com/MZMMSEC/AIGFD_BLO Area: Face Understanding / AI-Generated Content Detection Keywords: AI-generated face detection, bi-level optimization, self-supervised learning, EXIF metadata, anomaly detection

TL;DR

This paper proposes BLADES, a method that employs bi-level optimization to explicitly align self-supervised pretraining with the AI-generated face detection objective. The inner loop optimizes a visual encoder on pretext tasks including EXIF classification/ranking and face manipulation detection, while the outer loop optimizes task weights to improve performance on a proxy detection task, enabling cross-generator generalization without relying on any synthetic face data.

Background & Motivation

Faces synthesized by modern generative models (GANs, diffusion models) have reached a photorealistic level, creating an urgent need for robust detection methods. Core limitations of existing approaches: - Model-dependent detectors (supervised learning) tend to overfit generator-specific artifacts seen during training and fail to generalize to emerging generative techniques. - Model-agnostic detectors (relying on physiological inconsistencies such as pupil shape and corneal highlights as handcrafted features) may miss subtle statistical differences introduced by advanced generators. - Self-supervised feature methods exhibit better cross-generator performance but are still suboptimal, as their pretext tasks are not explicitly designed for the detection objective.

The key insight of BLADES is that conventional self-supervised learning pipelines do not explicitly align pretext losses with the downstream task, resulting in insufficiently specialized representations. Bi-level optimization provides a principled, mathematically grounded way to end-to-end steer self-supervised pretraining toward the downstream objective.

Method

Overall Architecture

The system consists of two stages: (1) bi-level optimization-driven self-supervised pretraining; and (2) detection inference with the encoder frozen. The pretraining stage adopts a joint image-text embedding architecture (CLIP-style, ResNet-50 visual encoder + GPT-2 text encoder) and is trained exclusively on real face photographs.

Key Designs

1. Bi-Level Optimization Formulation:

   • Inner Loop: On training set \(\mathcal{B}_{tr}\), optimize encoder parameters \(\boldsymbol{\theta}\) using a weighted combination of pretext task losses: \(\boldsymbol{\theta}^{\star} = \arg\min_{\boldsymbol{\theta}} \sum_{\boldsymbol{x} \in \mathcal{B}_{tr}} \sum_{i=1}^{K} \lambda_i \ell_i(\boldsymbol{x}; \boldsymbol{\theta})\)
   • Outer Loop: On validation set \(\mathcal{B}_{val}\), optimize task weights \(\boldsymbol{\lambda}\) to minimize the proxy detection task loss: \(\min_{\boldsymbol{\lambda}} \sum_{\boldsymbol{x} \in \mathcal{B}_{val}} \ell_1(\boldsymbol{x}; \boldsymbol{\theta}^{\star})\)
   • \(\boldsymbol{\theta}\) and \(\boldsymbol{\lambda}\) are updated alternately, automatically prioritizing pretext tasks most beneficial to downstream detection.

2. Four Pretext Task Designs:

   • Coarse-Grained Face Manipulation Detection (proxy main task \(\ell_1\)): Manipulated faces are generated via local flipping and global affine transformations; a fidelity loss trains the encoder to distinguish "manipulated" from "photographic" faces.
   • Categorical EXIF Tag Classification (\(\ell_i\)): Predicts categorical EXIF tags such as white balance mode and flash settings, using a focal fidelity loss to handle long-tail distributions.
   • Ordinal EXIF Tag Ranking (\(\ell_j\)): Discretizes numerical EXIF tags (ISO, aperture, etc.) into three levels—low/medium/high—and performs pairwise ranking via the Thurstone model.
   • Fine-Grained Face Manipulation Detection (\(\ell_m\)): Identifies specific manipulated facial regions (eyes, mouth, nose) and outputs region-level manipulation probabilities via sigmoid activations.

3. Detection Inference Strategies:

   • One-Class Anomaly Detection (BLADES-OC): Fits a 10-component GMM to real face features; at test time, samples whose likelihood falls below the 5th percentile threshold are classified as AI-generated.
   • Binary Classification Detection (BLADES-BC): Trains a lightweight two-layer perceptron (768→1536→2), augmenting AI-generated face samples with low-likelihood real faces as pseudo-outliers.

Loss & Training

• All tasks uniformly adopt a fidelity-based loss \(\ell = 1 - \sum \sqrt{p \cdot \hat{p}}\) (Bhattacharyya coefficient).
• A focal variant is applied to categorical EXIF tasks to address class imbalance, with \(\gamma=2\).
• Inner loop learning rate \(\alpha=10^{-5}\) (AdamW, cosine annealing); outer loop learning rate \(\beta=3 \times 10^{-4}\) (Adam).
• Pretraining for 20 epochs, batch size 48, input resolution \(224 \times 224\).
• Training data: 130K face photographs with EXIF metadata from the FDF dataset.

Key Experimental Results

Main Results (Cross-Generator Detection Accuracy %)

Method	StyleGAN2	VQGAN	LDM	DDIM	SDv2.1	Midjourney	SDXL	Avg.
CNND	50.61	99.89	53.07	56.55	50.51	51.66	54.49	58.41
FatFormer	98.91	98.30	97.82	95.63	68.88	88.20	88.08	89.68
Zou25	76.88	74.59	93.83	93.63	78.62	91.29	91.71	86.63
BLADES-OC	76.75	76.78	93.63	96.05	80.70	92.79	94.48	88.01
BLADES-BC	94.22	97.24	96.95	94.33	74.83	95.19	93.84	91.86

Ablation Study (Feature Separability AUC %)

Method	StyleGAN2	LDM	SDv2.1	FreeDoM	SDXL	Avg.
CLIP	33.99	55.49	90.15	85.39	93.66	76.13
FaRL	34.35	47.26	95.24	79.91	94.61	75.12
EAL	69.71	85.81	74.21	97.92	93.04	84.12
Zou25	85.69	98.66	88.29	99.79	97.23	93.43
BLADES-OC	87.89	98.24	91.72	99.92	98.36	95.05

Sensitivity Analysis

Method	TNR↑	FPR↓	TPR↑	FNR↓	F-score↑
LGrad	99.98	0.02	44.97	55.03	0.60
FatFormer	97.61	2.38	81.74	18.26	0.87
BLADES-BC	94.64	5.36	88.97	11.03	0.91

Key Findings

• BLADES-BC achieves an average accuracy of 91.86%, substantially outperforming all competing methods.
• BLADES-OC, trained solely on real faces, surpasses most supervised methods, validating the effectiveness of the bi-level optimization alignment strategy.
• Prior methods generally exhibit high specificity (TNR > 97%) but low sensitivity (TPR < 82%); BLADES-BC achieves a balanced trade-off between the two.
• Diffusion-model-based detection methods (e.g., DIRE, AEROBLADE) also generalize poorly within the same generative family, demonstrating that reliance on model-specific cues is inherently fragile.

Highlights & Insights

• The application of bi-level optimization is particularly elegant: the outer loop indirectly optimizes the true objective (AI-generated face detection) via a proxy task (manipulation detection), requiring no synthetic face training data whatsoever.
• The use of EXIF metadata as a self-supervised signal is novel—categorical tags are treated as classification targets while ordinal tags are used for ranking, fully exploiting the structural information in the metadata.
• The joint embedding architecture unifies heterogeneous pretext tasks within a single framework through text templates.

Limitations & Future Work

• Pretraining relies on face photographs with EXIF metadata; a large proportion of images on social media have their EXIF information stripped.
• Detection capability for GAN-generated images under the one-class setting remains relatively limited (StyleGAN2: 76.75% only).
• The binary classification setting still requires a small amount of synthetic data to fine-tune the classification head.

Related Work & Insights

• The idea of using bi-level optimization for task weight learning is generalizable to other multi-task self-supervised settings.
• EXIF metadata signals may be applicable to broader image provenance and integrity verification tasks.
• The fidelity loss (based on the Bhattacharyya coefficient) as an alternative to standard classification losses merits further attention.

Rating

• Novelty: ⭐⭐⭐⭐⭐ The idea of using bi-level optimization to align self-supervised pretraining with the detection objective is highly original, and the EXIF-based task design is comprehensive.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ Evaluations cover 9 generators, cross-dataset settings, feature separability, and sensitivity analysis—extremely thorough.
• Writing Quality: ⭐⭐⭐⭐ Problem formulation is clear and mathematical derivations are rigorous.
• Value: ⭐⭐⭐⭐ Addresses the generalization challenge in AI-generated content detection with significant application value for deepfake governance.

Related Papers

• [ICCV 2025] LVFace: Progressive Cluster Optimization for Large Vision Models in Face Recognition
• [ICCV 2025] Mitigating Object Hallucinations via Sentence-Level Early Intervention
• [CVPR 2026] RegFormer: Transferable Relational Grounding for Efficient Weakly-Supervised HOI Detection
• [ICCV 2025] CleanPose: Category-Level Object Pose Estimation via Causal Learning and Knowledge Distillation
• [ICCV 2025] Weakly Supervised Visible-Infrared Person Re-Identification via Heterogeneous Expert Collaborative Consistency Learning