FakeSpotter: A Simple Baseline for Spotting AI-Synthesized Fake Faces

Run Wang, Lei Ma, Felix Juefei-Xu, Xiaofei Xie, Jian Wang, Yang Liu

Introduction

The rise of Generative Adversarial Networks (GANs) has enabled the creation of hyper-realistic synthetic faces, ranging from entirely new identities to subtle expression edits and identity swaps (DeepFakes). While these technologies offer creative potential, they also present significant risks regarding misinformation and privacy. Existing detection methods often struggle with real-world conditions, such as image compression or resizing, which strip away the pixel-level artifacts these detectors rely on.

The Challenge of Robust Detection

Traditional forensic tools and deep-learning detectors typically analyze raw pixels or shallow features. However, when a fake image is uploaded to social media, it undergoes various perturbations—compression, blurring, or resizing—that often render these detectors ineffective. To address this, the authors of "FakeSpotter" propose a method that looks deeper than the surface pixels, focusing instead on the internal logic of deep neural networks.

Examples of the four types of facial manipulation addressed by FakeSpotter: Entire Synthesis, Attribute Editing, Expression Manipulation, and DeepFakes.

Methodology: Monitoring Neuron Behavior

FakeSpotter introduces a novel approach by monitoring the layer-by-layer neuron activation patterns within a pre-trained Face Recognition (FR) system, specifically a VGG-Face model with a ResNet-50 backbone. The core idea is that real and fake faces trigger different internal representations in a network trained to recognize human features.

The process involves:

  1. Mean Neuron Coverage (MNC): Establishing layer-specific activation thresholds based on average neuron outputs from training data.
  2. Feature Extraction: Passing an image through the FR system and counting how many neurons in each layer exceed the MNC threshold.
  3. Classification: Concatenating these counts into a feature vector, which is then processed by a simple 5-layer fully connected network to determine if the face is real or fake.

By abstracting the image into high-level neuron activation counts, FakeSpotter becomes significantly more robust to low-level pixel degradations.

Experimental Results

The researchers evaluated FakeSpotter against four categories of facial manipulation: Entire Synthesis (e.g., PGGAN, StyleGAN2), Attribute Editing (e.g., StarGAN, STGAN), Expression Manipulation, and DeepFakes (e.g., FaceForensics++, Celeb-DF).

FakeSpotter achieved an average detection accuracy exceeding 90%. Notably, it demonstrated exceptional robustness; when images were subjected to compression, noise, or blur, the system's performance (AUC) dropped by less than 3.77% on average. On the challenging Celeb-DF(v2) benchmark, FakeSpotter outperformed 13 existing state-of-the-art video detection methods.

Conclusion

FakeSpotter provides a robust baseline for synthetic face detection by leveraging the internal "knowledge" of existing face recognition systems. While it faces challenges with audio-only manipulations and potential adversarial attacks, its ability to maintain high accuracy under common image perturbations marks a significant step forward in practical digital forensics. Future work aims to expand this neuron-coverage methodology to multimodal detection, combining facial and vocal analysis.

Summary updated

Abstract

In recent years, generative adversarial networks (GANs) and its variants have achieved unprecedented success in image synthesis. They are widely adopted in synthesizing facial images which brings potential security concerns to humans as the fakes spread and fuel the misinformation. However, robust detectors of these AI-synthesized fake faces are still in their infancy and are not ready to fully tackle this emerging challenge. In this work, we propose a novel approach, named FakeSpotter, based on monitoring neuron behaviors to spot AI-synthesized fake faces. The studies on neuron coverage and interactions have successfully shown that they can be served as testing criteria for deep learning systems, especially under the settings of being exposed to adversarial attacks. Here, we conjecture that monitoring neuron behavior can also serve as an asset in detecting fake faces since layer-by-layer neuron activation patterns may capture more subtle features that are important for the fake detector. Experimental results on detecting four types of fake faces synthesized with the state-of-the-art GANs and evading four perturbation attacks show the effectiveness and robustness of our approach.
BibTeX
@inproceedings{wang2020fakespotter,
  author = {Run Wang and Lei Ma and Felix Juefei-Xu and Xiaofei Xie and Jian Wang and Yang Liu},
  title = {{FakeSpotter: A Simple Baseline for Spotting AI-Synthesized Fake Faces}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.24963/ijcai.2020/476},
}
Figure