Toward the Creation and Obstruction of DeepFakes

Yuezun Li, Pu Sun, Honggang Qi, Siwei Lyu, Pu Sun 0001Handbook of Digital Face Manipulation and Detection 2022 · DOI · DBLP

Abstract: AI-synthesized face-swapping videos, commonly known as DeepFakes, is an emerging problem threatening the trustworthiness of online information. The need to develop and evaluate DeepFake detection algorithms calls for large-scale datasets. However, current DeepFake datasets suffer from low visual quality and do not resemble DeepFake videos circulated on the Internet. We present a new large-scale challenging DeepFake video dataset, Celeb-DF, which contains 5,639 high-quality DeepFake videos of celebrities generated using an improved synthesis process. We conduct a comprehensive evaluation of DeepFake detection methods and datasets to demonstrate the escalated level of challenges posed by Celeb-DF. Then we introduce Landmark Breaker, the first dedicated method to disrupt facial landmark extraction, and apply it to the obstruction of the generation of DeepFake videos. The experiments are conducted on three state-of-the-art facial landmark extractors using our Celeb-DF dataset.

Introduction: The Evaluation and Reactive Gaps

DeepFake research currently faces a significant "lab vs. reality" mismatch. Most existing datasets, such as UADFV or FaceForensics++, contain obvious visual artifacts—blurry faces, visible splicing boundaries, and jittery movements—that make them relatively easy for automated detection systems to identify. However, real-world fakes are becoming increasingly sophisticated, rendering current detectors ineffective.

The authors identify two critical gaps in the field:

  1. The Evaluation Gap: Detectors trained on low-quality datasets fail when confronted with high-quality, realistic fakes.
  2. The Reactive Gap: Most research focuses on detecting fakes after they are disseminated. There is a lack of proactive methods to prevent fakes from being generated in the first place.
Visual artifacts like visible splicing boundaries in early DeepFake datasets.

Celeb-DF: Bridging the Quality Gap

To address the evaluation gap, the authors present Celeb-DF, a large-scale dataset containing 5,639 high-quality DeepFake videos of celebrities. Unlike previous benchmarks, Celeb-DF utilizes an improved synthesis process to minimize visual artifacts. The synthesis process was refined through several key technical improvements:

  • Resolution Enhancement: Synthesized face resolution was increased from standard $64 \times 64$ or $128 \times 128$ to $256 \times 256$.
  • Color Correction: Implementation of color transfer algorithms ensures the donor face's skin tone matches the target environment.
  • Seamless Masking: Landmark-based smooth masks replace simple rectangular masks to hide splicing boundaries.
  • Temporal Smoothing: Kalman filtering applied to facial landmarks reduces the "flickering" common in older generations of fakes.
Comparison of DeepFake quality across generations (rows 1-3) and the disruptive effect of Landmark Breaker (row 4).

Landmark Breaker: A Proactive Defense

DeepFake generation relies heavily on facial landmarks (key points on eyes, nose, and mouth) to align the donor face onto the target. The authors introduce Landmark Breaker, the first dedicated method to disrupt this process.

Landmark Breaker is a "white-box" adversarial attack that adds imperceptible noise to an image, misleading the neural networks responsible for landmark extraction. By targeting the "heat-maps" (probability maps) produced by these extractors, Landmark Breaker causes the alignment step to fail, resulting in garbled or poorly placed face swaps. This effectively sabotages the entire synthesis pipeline at its most vulnerable bottleneck.

Key Results and Insights

The introduction of Celeb-DF reveals that current state-of-the-art detectors are often "lazy," relying on simple signal-level artifacts rather than deep semantic inconsistencies. When these artifacts are removed in Celeb-DF, detector performance drops precipitously:

  • Detection Difficulty: On older datasets like UADFV, detectors achieved near-perfect accuracy (up to 97.7% AUC). On Celeb-DF, performance plummeted to an average of 56.9% AUC—barely better than random chance.
  • Obstruction Effectiveness: Landmark Breaker successfully disrupted top-tier landmark extractors (FAN, HRNet, and AVS-SAN), increasing the Normalized Mean Error (NME) from a baseline of 0.03 to 0.87.
  • Robustness: The adversarial perturbations remained effective even after JPEG compression or video re-encoding, which typically degrades such noise.

Conclusion and Future Work

While Landmark Breaker is highly effective in white-box scenarios, its transferability to "black-box" settings remains a challenge. Future research should focus on developing "universal" adversarial perturbations capable of disrupting any landmark extractor architecture. Such advancements would provide a robust "digital armor" for individuals seeking to protect their visual identity from unauthorized manipulation.

Summary updated

BibTeX
@inproceedings{li2022toward,
  author = {Yuezun Li and Pu Sun and Honggang Qi and Siwei Lyu and Pu Sun},
  title = {{Toward the Creation and Obstruction of DeepFakes}},
  booktitle = {Handbook of Digital Face Manipulation and Detection},
  year = {2022},
  doi = {10.1007/978-3-030-87664-7\_4},
}
Figure