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

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

TL;DR

This paper introduces FakeSpotter, a novel method for detecting AI-generated fake faces by analyzing the internal "neuron behaviors" of deep face recognition systems. By monitoring how neurons activate layer-by-layer, FakeSpotter captures subtle features that are difficult for current detectors to discern, offering a robust and effective baseline for identifying sophisticated fake imagery.

What Problem Does This Paper Address?

The rapid advancement of Generative Adversarial Networks (GANs) has led to the creation of highly realistic synthetic facial images that are increasingly difficult for humans to distinguish from real ones. While these technologies offer creative applications, they also pose significant security and privacy risks, enabling the spread of misinformation and the creation of malicious content. Existing methods for detecting these AI-synthesized faces are still in their early stages and struggle to keep pace with the evolving capabilities of GANs, particularly when faced with various image manipulations and adversarial attacks. This paper aims to bridge this gap by proposing a more robust and effective detection mechanism.

Methods

The core of FakeSpotter lies in its innovative approach to analyzing the internal workings of deep face recognition (FR) models. Instead of relying on raw pixel data or traditional image features, FakeSpotter monitors the neuron activation patterns within a pre-trained deep FR network. The key idea is that as an image passes through the layers of a deep network, the activation of neurons at each layer captures increasingly abstract and subtle features.

The authors introduce a metric called Mean Neuron Coverage (MNC). This metric is used to determine a dynamic threshold for neuron activation in each layer of the FR model. By calculating the average neuron output for a given layer across a set of training inputs (both real and fake faces), they establish a layer-specific threshold. A neuron is considered "activated" if its output exceeds this threshold.

These activation counts from each layer are then aggregated to form a feature vector for each input face. This feature vector, representing the "neuron behavior," is fed into a simple, shallow binary classifier (a few fully-connected layers) that is trained to distinguish between real and fake faces. This approach is designed to be robust because it focuses on the internal representations learned by the FR model, which are hypothesized to be less susceptible to direct pixel-level perturbations compared to raw image data.

Datasets

The paper utilizes several datasets for training and evaluation:

  • Real Faces: CelebA and Flicker-Faces-HQ (FFHQ) were used for their diversity.
  • Fake Faces:
    • Entire Synthesis: PGGAN and StyleGAN2 (from NVIDIA) were used to generate non-existent faces.
  • Attribute Editing: StarGAN (for hair color) and STGAN (for eyeglasses) were used.
  • Expression Manipulation: StyleGAN and STGAN were used to manipulate facial expressions (e.g., smiles).
  • DeepFake: Data from FaceForensics++ and Celeb-DF were used, which involve face swapping.
  • Public Datasets: The Celeb-DF dataset (specifically Celeb-DF(v2) for video detection) and the DeepFake Detection Challenge (DFDC) dataset were used for evaluation and comparison.

The authors emphasize that they generated many of the fake faces themselves using state-of-the-art GANs to ensure high quality and diversity.

Key Insights and Ideas

  • Neuron Behavior as a Signature: The most significant insight is that the layer-by-layer activation patterns of neurons in a deep face recognition system can serve as a powerful indicator of whether a face is real or synthesized. This moves beyond analyzing superficial artifacts to understanding the internal "understanding" of the network.
  • Robustness to Perturbations: The neuron coverage approach appears inherently more robust to common image transformations (like compression, blur, resizing, and noise) than methods relying on raw pixel data. This is because the internal representations are less directly affected by these changes.
  • Simplicity and Effectiveness: Despite the complex nature of GAN-generated fakes, FakeSpotter uses a relatively simple binary classifier on top of the extracted neuron features, demonstrating that sophisticated detection doesn't necessarily require overly complex models.
  • GAN Evolution Challenge: The paper highlights that as GANs become more advanced (e.g., StyleGAN2), detecting their outputs becomes harder, indicating an ongoing arms race in this field.

Key Results

  • High Detection Accuracy: FakeSpotter achieves an average detection accuracy of over 90% on four types of high-quality fake faces synthesized by SOTA GANs.
  • Robustness: The performance (measured by AUC) degrades by less than 3.77% across five different intensities of four common perturbation attacks (compression, blur, resizing, noise).
  • Competitive Performance: FakeSpotter outperforms prior work like AutoGAN in detecting fake faces.
  • Video Detection: On the Celeb-DF(v2) dataset for fake video detection, FakeSpotter achieved an AUC score of 66.8%, outperforming other existing methods.

Limitations and Future Work

  • DFDC Dataset Performance: FakeSpotter's performance on the DFDC dataset was not as ideal as on other datasets. The authors attribute this to DFDC involving face and voice swaps, while their method only considers facial images.
  • Adversarial Attacks: The paper acknowledges that attackers could potentially introduce adversarial noise to evade detection, and there's a trade-off between generating imperceptible fakes and evading detection.
  • Future Directions:
    • Developing methods to handle multi-modal fakes (e.g., combining face and voice manipulation).
  • Investigating the interplay between FakeSpotter and novel fake localization methods.
  • Exploring the application of the neuron coverage approach to other domains, such as fake speech detection.
  • The need for public databases with diverse, high-quality fake faces produced by the latest GANs.

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 Felix Juefei-Xu and Lei Ma and Xiaofei Xie and Yihao Huang and Jian Wang and Yang Liu},
  title = {{FakeSpotter: A Simple yet Robust Baseline for Spotting AI-Synthesized Fake Faces}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.24963/ijcai.2020/476},
}
Figure