Abstract
As DeepFake technology advances toward photorealism, the need for robust detection systems becomes critical. This paper introduces a novel detection pipeline that leverages Eulerian motion magnification to expose subtle temporal inconsistencies in forged videos, particularly those subject to heavy compression. Additionally, we address the prevalent issue of demographic bias in facial forgery detection. By introducing the Indian Forgery Dataset and conducting a systematic study on training data composition, we demonstrate that achieving a "critical mass" of diversity—specifically a 30% representation of under-represented groups—is sufficient to significantly mitigate the generalization gap in detection accuracy.
Introduction
The proliferation of DeepFakes poses significant societal risks, from misinformation to personal harassment. Current detection methods often fail in real-world scenarios due to two primary factors: high video compression, which erases fine spatial artifacts, and demographic bias, resulting from training sets dominated by Caucasian subjects. This research proposes a motion-centric detection approach and evaluates the impact of ethnic diversity on model fairness.
Methodology: Motion Magnification and 3D-SE-ResNet
Traditional detectors often rely on spatial artifacts like skin texture or eye reflections. However, video compression (e.g., H.264) typically smooths these details. Our approach shifts focus to temporal patterns using Eulerian Motion Magnification. This technique amplifies pixel-level changes over time, making the jittery "stitching" artifacts of DeepFakes visible to neural networks.
The core architecture is a modified 3D-SE-ResNet. By incorporating dense skip connections between residual blocks, the model efficiently captures multi-scale features. The network processes a concatenated input of the magnified video clip and a "difference map" (the delta between magnified and original frames), allowing it to learn both temporal context and localized manipulation cues.
Addressing Demographic Bias: The Indian Forgery Dataset
To combat the "generalization gap" where models underperform on non-Caucasian faces, we introduced the Indian Forgery Dataset. This dataset contains 200 pristine and 248 forged videos (created via FSGAN) of Indian subjects, featuring 149 unique identities.
We conducted a bias study by varying the "Merging Fraction" ($\alpha$), representing the proportion of Indian vs. Caucasian samples in the training set. We measured the performance gap between the majority and minority groups to determine the optimal data balance for fairness.
Our findings suggest a "30% Rule" for Fairness: once the under-represented group reaches approximately 30% of the training data ($\alpha = 0.3$), the performance gap narrows significantly. This indicates that perfectly balanced datasets are not strictly necessary to achieve equitable detection performance, provided a critical threshold of diversity is met.
Experimental Results
The proposed method was evaluated against standard benchmarks, including FaceForensics++ and Celeb-DF. Our 3D-SE-ResNet consistently outperformed established models like MesoNet and XceptionNet, particularly on high-compression videos where spatial detectors typically degrade.
Ablation studies confirmed that the integration of dense connections and motion magnification provided incremental gains in accuracy across all tested datasets (DF, F2F, FS, NT, FSh), validating the multi-component approach.
Conclusion and Future Work
By prioritizing temporal motion over spatial appearance, our method offers a more resilient solution to compressed DeepFake detection. Furthermore, our work highlights the importance of ethnic diversity in training data, providing a practical guideline for building fairer AI systems. Future research will expand to a wider range of ethnicities and evaluate the system's performance against emerging generative techniques like StyleGAN-based swaps.
Summary updated