Protecting President Zelenskyy against Deep Fakes

Matyás Bohácek, Hany FaridarXiv 2022 · OpenReview

TL;DR

This paper introduces a novel identity-based deepfake detection model that combines facial and gestural mannerisms to distinguish authentic videos of Ukrainian President Zelenskyy from sophisticated deepfakes. The model achieves high accuracy, demonstrating the effectiveness of a personalized approach to combating disinformation, especially in critical geopolitical situations.

What Problem Does This Paper Address?

The paper tackles the growing threat of deepfake technology, particularly its use in disinformation campaigns during conflicts. Specifically, it addresses the challenge of detecting deepfake videos that impersonate public figures, using the example of a deepfake of President Zelenskyy admitting defeat during the 2022 Russian invasion of Ukraine. While general deepfake detection models exist, they often struggle with generalization and can be vulnerable to sophisticated attacks. The authors aim to create a more robust and reliable detection method by focusing on the unique behavioral characteristics of a specific individual.

Methods

The core of the proposed solution is an identity-based behavioral model that captures distinctive speaking styles. This model is built upon two main components:

  • Facial Mannerisms: Leveraging the work of Agarwal et al. (2019), the system extracts facial landmark positions, facial action units (AUs), head pose, and eye gaze from video frames. The model specifically focuses on 16 AUs, head rotation (x and z axes), mouth width, and lip vertical distance, totaling 20 facial features.
  • Gestural Mannerisms: This is a novel addition to the facial-based approach. The system uses Blazepose to estimate arm and hand positions (shoulders, elbows, wrists). These coordinates are then normalized into a speaker-centric action plane, resulting in 12 gestural features.

These facial and gestural features are then combined to create a person-specific behavioral model. A one-class, non-linear Support Vector Machine (SVM) is trained as a novelty detection model. This SVM learns to identify "normal" behavior from authentic videos of the target individual (Zelenskyy) and flags deviations as potential deepfakes or imposters.

Datasets

  • Training Data: 506 minutes of authentic video of President Zelenskyy from four different contexts (public address, press briefing, bunker, armchair).
  • Decoy Data:
    • 57 minutes of interview-style videos of seven world leaders (Jacinda Ardern, Joe Biden, Kamala Harris, Boris Johnson, Wladimir Klitschko, Angela Merkel, and Vladimir Putin).
  • 50 minutes of video across 27 distinct individuals from the FaceForensics++ dataset.
  • 3 minutes of commissioned lip-sync deepfakes.
  • 1 minute of an "in-the-wild" deepfake video.

The paper does not introduce a new benchmark dataset but uses existing datasets and custom-generated deepfakes for evaluation.

Key Insights and Ideas

  • Identity-Specific Detection is Crucial: The paper strongly advocates for identity-specific models, arguing that while general deepfake detectors are improving, they still lag behind specialized approaches, especially when dealing with high-quality fakes.
  • Gestures Matter: The inclusion of gestural features significantly enhances detection accuracy. This is attributed to the difficulty in perfectly synthesizing realistic hand movements in deepfakes, making them a valuable signal for distinguishing real from fake.
  • Head Rotation is Highly Correlated: For President Zelenskyy, head rotation (e.g., nodding) was found to be highly correlated with his hand movements, underscoring the interplay between facial and gestural cues.
  • Combined Features are Key: Neither facial nor gestural features alone are sufficient for consistently detecting deepfakes of Zelenskyy. The combination of both is essential for achieving high accuracy.

Key Results

  • The combined facial and gestural model achieved a classification accuracy of 100.0% against real Zelenskyy videos and 97.1% against deep-fake Zelenskyy videos when evaluated against a dataset of seven world leaders and the FaceForensics++ dataset.
  • When compared to the best-performing DFDC model (which is a generic detector), the proposed identity-specific model significantly outperforms it on non-Zelenskyy data, achieving accuracies like 83.9% on in-the-wild deepfakes versus the DFDC model's 1.7%.
  • An ablation study showed that while a significant fraction of features are informative, the full set of 496 pairwise facial and gestural correlations is needed for near-perfect detection (99.88% accuracy).
  • The top 20 most discriminative correlation features highlighted the importance of head pose and specific facial action units in conjunction with limb positions.

Limitations and Future Work

  • Identity Specificity: The model is trained for a single individual (Zelenskyy). While this is effective for protecting a specific high-profile figure, it requires retraining for each new target, which can be impractical if hours of authentic video are not available.
  • Data Requirements: The identity-based approach necessitates a substantial amount of authentic video data for training, which might not be available for all individuals.
  • Future Work: The authors suggest that this approach could be extended to protect other world leaders. They also posit that the combination of facial and gestural signals will remain reliable for several years due to the inherent difficulty in synthesizing realistic hand gestures.

Summary updated

Abstract

The 2022 Russian invasion of Ukraine is being fought on two fronts: a brutal ground war and a duplicitous disinformation campaign designed to conceal and justify Russia's actions. This campaign includes at least one example of a deep-fake video purportedly showing Ukrainian President Zelenskyy admitting defeat and surrendering. In anticipation of future attacks of this form, we describe a facial and gestural behavioral model that captures distinctive characteristics of Zelenskyy's speaking style. Trained on over eight hours of authentic video from four different settings, we show that this behavioral model can distinguish Zelenskyy from deep-fake imposters.This model can play an important role -- particularly during the fog of war -- in distinguishing the real from the fake.
BibTeX
@misc{bohacek2022protecting,
  author = {Matyás Bohácek and Hany Farid},
  title = {{Protecting President Zelenskyy against Deep Fakes}},
  year = {2022},
  eprint = {2206.12043},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2206.12043},
}
Figure