Exploring the Impact of Moire Pattern on Deepfake Detectors

Razaib Tariq, Shahroz Tariq, Simon S. Woo

Abstract

While state-of-the-art deepfake detectors demonstrate high accuracy on digital source files, their performance in real-world "re-capture" scenarios—where a video is filmed from a screen—remains largely unexplored. This paper investigates how Moiré patterns, resulting from the interference between a camera's pixel grid and a digital display, act as a natural spoofing attack. Our empirical study reveals that these patterns cause detection accuracy to plummet from over 90% to an average of less than 68%, highlighting a critical vulnerability in current deepfake detection frameworks.

1. Introduction

The proliferation of deepfake technology has necessitated the development of robust detection models. Most existing research focuses on digital-to-digital detection, assuming the model has access to the original file. However, a common real-world "spoofing" method involves playing a deepfake on a monitor and recording it with a secondary device, such as a smartphone, before distribution. This physical re-capture introduces Moiré patterns—wavy, rainbow-like interference artifacts. While these artifacts are often visible to humans, they significantly confound the deep neural networks (DNNs) that rely on fine-grained texture and frequency analysis to identify GAN-generated content.

2. Methodology

To quantify the impact of Moiré patterns, we established a controlled re-capture pipeline to test the resilience of existing models.

2.1 Experimental Setup

The hardware configuration consisted of an LG monitor (1980x1080 resolution) and a Samsung S22 Plus smartphone. The smartphone was mounted on a tripod at a fixed distance of 30cm from the screen to ensure consistent capture quality across the dataset.

2.2 Dataset Transformation

We utilized two primary benchmark datasets to create our re-captured versions:

  • CelebDF: High-quality deepfake videos of celebrities.
  • FaceForensics++ (FF++): A comprehensive dataset featuring four manipulation types: DeepFake, Face2Face, FaceSwap, and Neural Textures.

From these, we generated a "camera-captured" dataset by re-recording a balanced subset of 316 videos, totaling 110,600 frames. This dataset contains the inherent Moiré patterns and environmental noise typical of physical-world spoofing.

2.3 Evaluated Models

We tested six state-of-the-art (SOTA) detectors: CoReD, CLRNet, Xception, BZNet, QAD, and ADD. These models represent the current standard in deepfake detection and were originally trained on "clean" digital datasets, which is the current industry standard for deployment.

3. Key Results and Analysis

The introduction of Moiré patterns resulted in a drastic reduction in model efficacy across all metrics, revealing a significant generalization gap.

3.1 Performance Degradation

The average F1-score for all tested models dropped significantly, with no model exceeding a 68% average accuracy on re-captured data. The Area Under the Receiver Operating Characteristic (AUROC) curve, a key measure of discriminative power, peaked at a mere 0.64 (CLRNet), suggesting that the models have very limited ability to distinguish between real and fake content in the presence of interference.

3.2 Model-Specific Findings

  • Most Robust: The ADD model proved the most resilient, reaching 86% accuracy on specific subsets, though its overall average remained significantly lower than its performance on clean data.
  • Least Robust: CoReD struggled the most, with accuracy falling as low as 47%—essentially performing worse than a coin flip.
  • Overfitting: Models like BZNet showed high variance, achieving 94% accuracy on specific subsets (FF++ DeepFake) but dropping to near-random levels on others. This suggests these models are learning dataset-specific quirks rather than generalizable features of facial manipulation.

4. Discussion: Moiré Patterns as Accidental Adversaries

Moiré patterns are not merely visual noise; they fundamentally alter the texture and frequency distribution of the video frames. Since many SOTA detectors are trained to identify subtle GAN-generated artifacts in the frequency domain, the high-frequency interference introduced by the screen-camera interaction effectively masks the "fakeness" of the video. This creates a significant vulnerability where a simple physical act (re-recording a screen) acts as a successful adversarial attack.

5. Conclusion and Future Work

Current deepfake detection technology is highly vulnerable to physical-world environmental factors. To bridge this gap, future research must focus on:

  • Physical-World Training: Incorporating re-captured data or synthetic Moiré augmentations during the training phase to improve robustness.
  • Hardware Diversity: Expanding research to include various screen types (OLED vs. LCD) and camera sensors to ensure broad applicability.
  • Defense Mechanisms: Developing novel preprocessing techniques to "de-moiré" videos or using data augmentation to help models ignore interference artifacts.

This study serves as a wake-up call for the community to move beyond digital-only benchmarks and address the complexities of physical-world spoofing attacks.

Summary updated

BibTeX
@inproceedings{tariq2024exploring,
  author = {Razaib Tariq and Shahroz Tariq and Simon S. Woo},
  title = {{Exploring the Impact of Moire Pattern on Deepfake Detectors}},
  booktitle = {ICIP},
  year = {2024},
  doi = {10.1109/icip51287.2024.10647902},
}
Figure