Open-World Deepfake Attribution via Confidence-Aware Asymmetric Learning

Haiyang Zheng1, Nan Pu1*, Wenjing Li2*, Teng Long1, Nicu Sebe1, Zhun Zhong2
1University of Trento, 2Hefei University of Technology
AAAI 2026

*Corresponding author
Detective Conan Analogy

OWDFA-CAL utilizes a Confidence-Aware Asymmetric Learning framework to act like a detective, identifying known criminals (Known Deepfakes) and uncovering mysterious new suspects (Novel Deepfakes) in the open world.

Task & Motivation

What is OWDFA?

Open-World Deepfake Attribution (OWDFA)

It acts like a detective. Unlike traditional Deepfake Detection which only performs binary classification (Real vs. Fake), OWDFA is distinct in the following aspects:

  • Source Tracing: Distinguish specifically which model created the forgery (e.g., FaceDancer, Diffusion, GANs).
  • The "Open World" Challenge: Unlabeled Data contains a complex mixture of Known and Novel types.
  • Goal: Classify knowns while discovering and clustering novels.
OWDFA Task Overview
Figure 1. The illustration of Open-World Deepfake Attribution.

Confidence Skew
Observation 1: Confidence Skew
Noise Proportion
Observation 2: Noise Proportion

Limitations of Existing Methods

Despite recent progress, current approaches suffer from two critical limitations:

1. Confidence Skew (The Bias)

Existing methods exhibit a severe confidence gap: high confidence for known forgeries but low for novel ones. This creates a "rich get richer" loop where novel classes are ignored.

2. Unrealistic Assumption on K

Most methods assume the number of novel forgery types (K) is known a priori. Our method (CAL) removes this assumption by estimating K dynamically.

Abstract

The proliferation of synthetic facial imagery has intensified the need for robust Open-World DeepFake Attribution (OWDFA), which aims to attribute both known and unknown forgeries using labeled data for known types and unlabeled data containing a mixture of known and novel types. However, existing methods face two critical limitations: 1) A confidence skew leading to unreliable pseudo-labels for novel forgeries; 2) An unrealistic assumption that the number of unknown forgery types is known a priori.

To address these challenges, we propose a Confidence-Aware Asymmetric Learning (CAL) framework. CAL consists of two key components: Confidence-Aware Consistency Regularization (CCR), which mitigates pseudo-label bias by dynamically scaling sample losses, and Asymmetric Confidence Reinforcement (ACR), which separately calibrates confidence for known and novel classes. Together, they form a mutually reinforcing loop that significantly improves performance.

Moreover, we introduce a Dynamic Prototype Pruning (DPP) strategy to automatically estimate the number of novel forgery types in a coarse-to-fine manner, removing unrealistic prior assumptions. Extensive experiments on the standard OW-DFA benchmark and a newly extended OWDFA-40 benchmark demonstrate that CAL consistently achieves new state-of-the-art performance.

Method: Confidence-Aware Asymmetric Learning

CAL Framework Pipeline

Core Components of CAL

Confidence-Aware Consistency Regularization (CCR)

Uses a threshold-free dynamic weighting strategy. It adaptively shifts training focus from high-confidence known samples to low-confidence novel ones, effectively mitigating the impact of noisy pseudo-labels to ensure stable learning.

Asymmetric Confidence Reinforcement (ACR)

Leverages dynamically adjusted asymmetric thresholds to independently select reliable pseudo-labels for known and novel classes, effectively bridging the confidence gap.

Dynamic Prototype Pruning (DPP)

Estimates the number of novel forgery types without prior knowledge in a coarse-to-fine manner:

High-confidence Identification Low-confidence Filtering Similarity-driven Merge

Experiments

1. The OWDFA-40 Benchmark

To simulate realistic open-world scenarios, we constructed the OWDFA-40 Benchmark, extending the original setup to include 40 advanced deepfake generation methods.

  • Diversity: Covers 5 categories including Face Swapping, Reenactment, Editing, Entire Face Synthesis, and Diffusion-based Generation.
  • Protocols: We designed three protocols with varying degrees of "openness" to rigorously evaluate robustness against unseen paradigms.
Overview of Deepfake Methods in OWDFA-40
Table 1. Overview of Deepfake Methods and Protocol Splits.

2. Comparison with State-of-the-Art

We compare CAL with SOTA methods (CPL, CDAL) and strong baselines from GCD and OWSSL settings.

Key Finding: CAL achieves a new state-of-the-art performance across all three protocols. Notably, it surpasses CPL by an average of 7.3% in Novel ACC, demonstrating superior ability to discover unknown forgery types.
Unknown K: Even without knowing the number of novel types (w/o KU), our method outperforms baselines that rely on this prior knowledge.

SOTA Comparison Results
Table 2. Quantitative results on the OW-DFA-40 benchmark. Best results are bolded.

3. Ablation Study

We conduct a step-by-step ablation to validate the effectiveness of our key components:

  • Base Loss: Foundational performance.
  • CCR: Boosts Novel ACC (+32.8%) by mitigating noise.
  • ACR: Bridges the confidence gap.
  • FFE: Improves generalized representation.
Ablation Study Table
Table 3. Effectiveness of Key Components.

4. Estimating Unknown K

Unlike GCD methods that rely on clustering accuracy of known classes to guess K, our Dynamic Prototype Pruning (DPP) strategy directly estimates the number of novel forgeries with much lower error.

Class Number Estimation Results
Table 4. Estimation Error Comparison.

Poster Presentation

Project Poster

BibTeX

@article{zheng2025open,
  title={Open-World Deepfake Attribution via Confidence-Aware Asymmetric Learning},
  author={Zheng, Haiyang and Pu, Nan and Li, Wenjing and Long, Teng and Sebe, Nicu and Zhong, Zhun},
  journal={arXiv preprint arXiv:2512.12667},
  year={2025}
}