Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FeatDistill: A Feature Distillation Enhanced Multi-Expert Ensemble Framework for Robust AI-generated Image Detection

Zhilin Tu, Kemou Li, Fengpeng Li, Jianwei Fei, Jiamin Zhang, Haiwei Wu

Abstract

The rapid iteration and widespread dissemination of deepfake technology have posed severe challenges to information security, making robust and generalizable detection of AI-generated forged images increasingly important. In this paper, we propose FeatDistill, an AI-generated image detection framework that integrates feature distillation with a multi-expert ensemble, developed for the NTIRE Challenge on Robust AI-Generated Image Detection in the Wild. The framework explicitly targets three practical bottlenecks in real-world forensics: degradation interference, insufficient feature representation, and limited generalization. Concretely, we build a four-backbone Vision Transformer (ViT) ensemble composed of CLIP and SigLIP variants to capture complementary forensic cues. To improve data coverage, we expand the training set and introduce comprehensive degradation modeling, which exposes the detector to diverse quality variations and synthesis artifacts commonly encountered in unconstrained scenarios. We further adopt a two-stage training paradigm: the model is first optimized with a standard binary classification objective, then refined by dense feature-level self-distillation for representation alignment. This design effectively mitigates overfitting and enhances semantic consistency of learned features. At inference time, the final prediction is obtained by averaging the probabilities from four independently trained experts, yielding stable and reliable decisions across unseen generators and complex degradations. Despite the ensemble design, the framework remains efficient, requiring only about 10 GB peak GPU memory. Extensive evaluations in the NTIRE challenge setting demonstrate that FeatDistill achieves strong robustness and generalization under diverse ``in-the-wild'' conditions, offering an effective and practical solution for real-world deepfake image detection.

FeatDistill: A Feature Distillation Enhanced Multi-Expert Ensemble Framework for Robust AI-generated Image Detection

Abstract

The rapid iteration and widespread dissemination of deepfake technology have posed severe challenges to information security, making robust and generalizable detection of AI-generated forged images increasingly important. In this paper, we propose FeatDistill, an AI-generated image detection framework that integrates feature distillation with a multi-expert ensemble, developed for the NTIRE Challenge on Robust AI-Generated Image Detection in the Wild. The framework explicitly targets three practical bottlenecks in real-world forensics: degradation interference, insufficient feature representation, and limited generalization. Concretely, we build a four-backbone Vision Transformer (ViT) ensemble composed of CLIP and SigLIP variants to capture complementary forensic cues. To improve data coverage, we expand the training set and introduce comprehensive degradation modeling, which exposes the detector to diverse quality variations and synthesis artifacts commonly encountered in unconstrained scenarios. We further adopt a two-stage training paradigm: the model is first optimized with a standard binary classification objective, then refined by dense feature-level self-distillation for representation alignment. This design effectively mitigates overfitting and enhances semantic consistency of learned features. At inference time, the final prediction is obtained by averaging the probabilities from four independently trained experts, yielding stable and reliable decisions across unseen generators and complex degradations. Despite the ensemble design, the framework remains efficient, requiring only about 10 GB peak GPU memory. Extensive evaluations in the NTIRE challenge setting demonstrate that FeatDistill achieves strong robustness and generalization under diverse ``in-the-wild'' conditions, offering an effective and practical solution for real-world deepfake image detection.
Paper Structure (20 sections, 5 equations, 2 figures, 2 tables)

This paper contains 20 sections, 5 equations, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Deepfake and Image Forgery Detection
  4. Detection in the Diffusion Era
  5. Generalization and Open-World Robustness
  6. Pre-trained Backbones and Distillation
  7. Competition Data
  8. Proposed Method: FeatDistill
  9. Architecture Overview
  10. Data Strategy and Degradation Modeling
  11. Two-Stage Training with Dense Supervision
  12. Multi-Expert Ensemble Inference
  13. Backbone Perspective
  14. Data Perspective
  15. Training Perspective
  16. ...and 5 more sections

Figures (2)

  • Figure 1: Illustration of the robust AI-generated image detection task in the wild. Real and AI-generated images are both exposed to complex perturbations encountered in practical media pipelines, including motion blur, Gaussian noise, JPEG compression, resizing, color adjustment, lens distortion, filtering, and screenshot/social-media reposting effects. These factors induce severe domain shifts and interactions with unseen generators. A practical detector should therefore produce calibrated probabilities and robust real/AI predictions under open-world deployment constraints.
  • Figure 2: Overview of the proposed FeatDistill framework. Our approach utilizes a multi-expert ensemble consisting of four high-capacity backbones. The training follows a two-stage paradigm: Stage 1 establishes a discriminative baseline, while Stage 2 employs dense feature-level self-distillation to enhance structural consistency. Robustness is further ensured through multi-source data expansion and a comprehensive 35-algorithm degradation library.