Abstract
Visual content has become the primary source of information, leading to an increase in video alterations. While some are simple, sophisticated alterations like reenactment-based DeepFakes are difficult to detect. Reenactment allows a source to change target expressions, creating photo-realistic results with significant social implications. This research proposes a learning-based algorithm using a multi-stream network that learns regional artifacts to provide robust performance across various compression levels. Using a custom loss function for balanced learning, the proposed network achieves state-of-the-art classification accuracy on the FaceForensics dataset, reaching 91.20% even under heavy compression.
Introduction: The Challenge of Facial Reenactment
As DeepFake technology advances, "facial reenactment" has emerged as a significant concern. Unlike simple identity swapping, reenactment allows a source actor to manipulate a target's expressions and speech in real-time. This capability can be used maliciously to make individuals appear to say or do things they never did.
A major hurdle for existing detection algorithms is video compression. While detectors often perform well on high-quality, uncompressed images, the heavy compression used by social media platforms like YouTube or WhatsApp "washes out" the subtle artifacts and noise patterns that these systems rely on. This research aims to create a robust detection system that maintains high performance across various compression levels.
Proposed Methodology: Multi-Stream Learning
The authors propose a multi-stream architecture based on ResNet-18 to capture both global and local inconsistencies.
1. Preprocessing and Regional Division
Faces are first detected using the S3FD algorithm. Each detected face is cropped and divided into a $2 \times 2$ grid, creating four local regions: the left eye, right eye, left cheek/mouth, and right cheek/mouth.
2. Multi-Stream Architecture
The network consists of five parallel ResNet-18 streams. One stream processes the full face, while the remaining four process the local regions. The outputs from these streams are concatenated into a 10-dimensional feature vector, which is then passed through a final fully connected layer for binary classification (Real vs. Altered).
3. Balanced Loss Function
To prevent the network from over-relying on the full-face stream, the authors introduced a custom balanced loss function ($L_{total}$). This function combines individual cross-entropy losses for each stream with a global fusion loss, forcing the regional streams to learn localized artifacts that might be overlooked in a global view.
Experimental Results
The model was evaluated using the FaceForensics dataset, which contains 1,004 videos manipulated via the Face2Face approach. Performance was tested across three H.264 compression levels: No-compression (QP 0), Easy-compression (QP 23), and Hard-compression (QP 40).
The proposed method outperformed existing benchmarks, particularly in the most challenging "hard compression" category, which simulates heavy social media processing.
| Method | No-Compression | Easy-Compression | Hard-Compression |
|---|---|---|---|
| MesoNet | 96.80% | 93.40% | 83.20% |
| XceptionNet | 99.93% | 98.13% | 87.81% |
| Proposed Approach | 99.96% | 99.10% | 91.20% |
Qualitative analysis using Class Activation Maps (CAMs) confirmed that the regional classifiers successfully focused on the eyes and mouth areas—regions where the Face2Face algorithm often leaves subtle artifacts during rapid movement.
Conclusion
By combining global context with localized "expert" streams, this multi-stream approach provides a robust defense against facial reenactment, even in low-quality video environments. Future work will focus on improving computational efficiency—perhaps through attention mechanisms—and testing the model's generalization across other types of generative facial manipulations.
Summary updated