TL;DR
This paper introduces a novel method for distinguishing AI-generated human motion videos from real ones by leveraging CLIP embeddings. Their approach is robust to common video manipulation techniques and achieves high accuracy, even on unseen AI models, paving the way for more reliable detection of synthetic media.
What Problem Does This Paper Address?
The rapid advancement of AI in generating realistic video content, particularly human motion, presents a dual-edged sword. While offering creative possibilities, it also opens doors for malicious applications like deepfakes, misinformation, and non-consensual content. Existing detection methods often struggle with "laundering" techniques such as resolution changes or compression, which can obscure low-level artifacts. This paper aims to bridge this gap by developing a robust method that can reliably distinguish real human motion from AI-generated content, even when subjected to such manipulations.
Methods
The core of the proposed solution lies in using multi-modal semantic embeddings, specifically variants of CLIP (Contrastive Language–Image Pre-training). Instead of relying on low-level visual artifacts, which are easily defeated by video processing, the authors leverage the semantic understanding captured by CLIP. They explore four types of embeddings:
- CLIP: A standard pre-trained CLIP model.
- SigLIP: A larger, sigmoid-loss-based variant of CLIP.
- JinaCLIP: A CLIP model trained with an emphasis on joint text-image and text-text alignment.
- FT-CLIP: A custom fine-tuned version of the baseline CLIP model, specifically trained for this task.
These embeddings are then used in conjunction with three classification strategies:
- Two-Class SVM: A standard Support Vector Machine trained to classify videos as either "real" or "fake" (bundling all AI models into one class).
- Multi-Class SVM: An SVM trained to classify videos into eight distinct classes: one for real videos and seven for each of the specific AI models used.
- Frame-to-Prompt: An unsupervised method that compares video frame embeddings directly against text embeddings of prompts like "real photo" vs. "fake photo," requiring no explicit training.
Datasets
The authors introduce a new benchmark dataset called DeepAction. This dataset comprises:
- AI-generated videos: 3,100 clips generated by seven different text-to-video AI models (BD AnimateDiff, CogVideoX-5B, Lumiere, RunwayML Gen3, Stable Diffusion Txt2Img+Img2Vid, Veo, and VideoPoet). These clips cover 100 distinct human actions and vary in length and resolution.
- Real videos: A curated set of 100 real videos from Pexels, matched to the human-action prompts used for the AI-generated videos.
The dataset is made publicly available under an academic license.
Key Insights and Ideas
- Semantic Embeddings are Key: The paper strongly suggests that relying on semantic features, as captured by models like CLIP, is far more robust to common video laundering techniques than low-level artifact detection.
- Fine-tuning Matters: While off-the-shelf CLIP models perform well, fine-tuning CLIP (FT-CLIP) on a relevant dataset significantly boosts detection accuracy, especially for distinguishing between specific AI models.
- Unsupervised Detection is Promising: The frame-to-prompt method, which requires no explicit training, offers a compelling alternative for deployment due to its inherent generalizability and simplicity.
- AI Models Have Distinct Signatures: Even when generating similar actions, different AI models leave subtly different "fingerprints" in their output, which can be captured by sophisticated embeddings.
Key Results
- High Accuracy: The best-performing model, FT-CLIP with a two-class SVM, achieved a video-level accuracy of 99.2% on the DeepAction dataset.
- Robustness to Laundering: The method showed remarkable resilience to resolution reduction (down to 50% of original) and compression (down to 10% of original bit rate), maintaining high accuracy.
- Generalizability: A leave-one-out analysis demonstrated that the SVM classifiers could generalize well to unseen AI models, with average video-level accuracy dropping only slightly from 96.9% to 95.2%.
- Non-Human Motion: The model also showed good generalization to non-human AI-generated content, correctly identifying 97.1% of videos as AI-generated, suggesting it captures general AI generation properties rather than just human motion specifics.
- Limitations with Face-Swap/Lip-Sync: The method struggled with face-swap and lip-sync deepfakes, achieving significantly lower accuracy, indicating that these types of manipulation, which often preserve more of the original video's structure, pose a different challenge.
Limitations and Future Work
- Face-Swap/Lip-Sync Deepfakes: The current approach does not generalize well to face-swap and lip-sync deepfakes, which are a significant concern. Future work could explore adapting the method or developing specialized techniques for these types of manipulations.
- Adversarial Attacks: While robust to standard laundering, the authors acknowledge the need to investigate resilience against more sophisticated adversarial attacks.
- Understanding Embeddings: The exact semantic properties that make these embeddings effective for AI detection are not fully understood and warrant further investigation.
- Ethical Considerations: The authors chose not to open-source their code to mitigate the risk of enabling the creation of more sophisticated deepfakes, highlighting the ethical tightrope in this research area.
Summary updated