Source-linked AI summary
Detecting Deep-Fake Videos from Appearance and Behavior
Shruti Agarwal, Tarek El-Gaaly, Hany Farid, Ser-Nam Lim
TL;DR
Face-swap deep fakes pose risks because realistic manipulated videos can support pornography, fraud, disinformation, and election disruption, while pixel-artifact detectors are vulnerable to laundering and poor generalization. The paper combines facial identity with learned temporal behavior to authenticate four-second clips against authentic references, and reports high effectiveness across five large datasets and previously unseen deep fakes. Its scope is limited because lip-sync fakes preserve facial identity and most behavior, making them difficult for this approach to classify.
Problem
Face-swap deep fakes require detectors less vulnerable to resizing and transcoding and better able to generalize beyond known synthesis techniques.
Method
The method combines a facial-recognition biometric with a CNN-learned behavioral biometric based on facial expressions and head movements.
Results
The combined biometrics are highly effective across five large datasets containing thousands of identities and generalize to previously unseen deep fakes and people.
Takeaways & Limitations
Matching facial and behavioral identities provides a forensic signal for detecting face-swap deep fakes across datasets and unseen identities.
Takeaways & Limitations
The approach will likely struggle with lip-sync deep fakes because facial identity and most behavior remain consistent with the depicted person.
Abstract
from arXiv · showhide
Synthetically-generated audios and videos -- so-called deep fakes -- continue to capture the imagination of the computer-graphics and computer-vision communities. At the same time, the democratization of access to technology that can create sophisticated manipulated video of anybody saying anything continues to be of concern because of its power to disrupt democratic elections, commit small to large-scale fraud, fuel dis-information campaigns, and create non-consensual pornography. We describe a biometric-based forensic technique for detecting face-swap deep fakes. This technique combines a static biometric based on facial recognition with a temporal, behavioral biometric based on facial expressions and head movements, where the behavioral embedding is learned using a CNN with a metric-learning objective function. We show the efficacy of this approach across several large-scale video datasets, as well as in-the-wild deep fakes.
1 Introduction
Deep fakes create serious risks because realistic manipulated media can enable abuse, fraud, disinformation, and election disruption. The paper addresses weaknesses in pixel-artifact detection with a biometric approach combining facial identity and temporal behavior.
- Motivation: Deep fakes can synthesize realistic audio, images, and videos, including videos depicting anyone saying or doing anything.The paper emphasizes that these capabilities are entertaining but can also be weaponized.
- Motivation: Deep fakes have been used for non-consensual pornography and may support fraud, election disruption, disinformation campaigns, and civil unrest.The paper identifies harms affecting women as well as broader political and social harms.
- Problem: Pixel-artifact detectors are vulnerable to transcoding and resizing and often struggle to generalize to new synthesis techniques.These limitations motivate a forensic technique based on higher-level biometric inconsistencies.
- Approach: The proposed detector combines facial identity with facial expressions and head movements to expose inconsistencies in face-swap videos.The behavioral embedding is learned with a CNN using a metric-learning objective, while facial identity uses standard face-recognition techniques.
- Evaluation: The experiments cover thousands of identities and include both large-scale datasets and in-the-wild deep fakes.The paper presents this evaluation as evidence for the efficacy of its approach.
- Contributions: The paper contributes a generic spatiotemporal behavior model, a combined appearance-and-behavior biometric detector, and evaluation across five large datasets.The contributions also include analysis of learned features and robustness across datasets, manipulations, and fake-video qualities.
2 Related Work
Prior deep-fake work spans multiple synthesis techniques and forensic detectors. The paper positions its method among low-level artifact detectors and high-level approaches that use semantic inconsistencies or identity-specific behavior.
- Deep-fake generation: Deep fake refers to synthetically generated images and videos, often of people, produced by techniques including FaceSwap, FS-GAN, Neural Textures, and Face2Face.The term covers multiple generation approaches rather than one specific synthesis method.
- Deep-fake generation: DeepFake FaceSwap replaces one person’s video identity with another’s using a GAN, while Faceswap-GAN uses an autoencoder with adversarial loss.These methods can generate convincing fakes but may require substantial training data.
- Deep-fake generation: Neural Textures supports lip-sync deep fakes by modifying a person’s mouth to match a new audio track.This extends earlier approaches that created lip-sync fakes for specific individuals.
- Deep-fake generation: Face2Face maps one person’s facial expressions and head movements onto another, whereas FaceSwap builds and aligns a 3-D facial model.These are traditional computer-graphics approaches rather than learning-based synthesis methods.
- Deep-fake detection: Low-level forensic methods detect synthesis artifacts, camera-metadata inconsistencies, or manipulated facial regions.Examples include CNN, Siamese-network, fully convolutional, and patch-based detectors.
- Deep-fake detection: Low-level detectors can be sensitive to resizing, transcoding, adversarial attacks, and novel datasets, whereas high-level methods tend to be more resilient.The paper uses this contrast to motivate semantically meaningful detection signals.
- Deep-fake detection: High-level detectors exploit semantic inconsistencies such as mismatched head pose between swapped and surrounding facial regions.These approaches use properties such as 3-D geometry that synthesis systems may find difficult to correct.
- Relation to prior work: The paper builds on behavior modeling by learning a CNN-based metric-learning biometric and pairing it with facial identity for four-second clips.Unlike earlier person-specific modeling, the approach targets identity consistency across reference and test videos.
3 Biometrics
The authentication system represents a video through static facial appearance and learned temporal behavior, then compares both against authentic identity references. A mismatch or weak facial match indicates a fake.
- Biometric signals: The system uses temporal behavioral and static facial biometrics based on expressions, head movements, and facial features.These measurements provide complementary signals for authentication.
- Behavior biometric: FAb-Net features encode head pose, facial landmarks, and facial expression, providing building blocks for spatiotemporal behavior.The paper stacks static features over time to represent behavior in video clips.
- Behavior biometric: The learned behavior embedding preserves facial movements and expressions while distinguishing identity-specific spatiotemporal behavior.This modifies identity-agnostic FAb-Net features through a learned low-dimensional mapping.
- Behavior biometric: The CNN maps 256 × t FAb-Net matrices into an embedding where cosine similarity is high for same-identity clips and low for different identities.The output is normalized to a unit sphere, enabling cosine-similarity comparison.
- Behavior biometric: Metric learning selects hard positive and negative samples, then applies soft weighting to emphasize informative examples during embedding training.The loss averages these weighted contributions over a mini-batch, with α, β, and λ as hyperparameters.
- Behavior biometric: Supervised identity-label training encourages the embedding space to cluster biometric signatures by identity.The model is trained on VoxCeleb2 using four-second, 100-frame clips and a ResNet-101 architecture.
- Authentication: For authentication, facial and Behavior-Net features are extracted from a test clip and matched by cosine similarity against authentic reference identities.The reference sets contain facial and behavioral feature matrices for each identity.
4 Results
Across five datasets and in-the-wild videos, the biometric detector was evaluated for accuracy, robustness, and feature behavior. Results were strong overall, while performance was weaker when face swaps failed to convincingly transfer source appearance.
- Overall detection: 95.5% average cross-over accuracy was achieved, with dataset accuracies ranging from 88.4% on DFDC-P to 99.1% on FF.The reported cross-over accuracies were 97.3% for WLDR, 99.1% for FF, 93.1% for DFD, 88.4% for DFDC-P, and 98.3% for CDF.
- Overall detection: At the fixed threshold τf = 0.86, detection accuracy ranged from 82.4% for DFDC-P to 98.9% for FF.These values correspond to the classification accuracies reported for the fixed facial VGG threshold.
- Failure analysis: DFDC-P performance was unusually low because many fakes failed to map the desired source appearance, leaving both behavior and appearance aligned with the target identity.The DFD dataset exhibited a similar source-to-target mapping problem, motivating a confidence threshold for facial similarity matching.
- In-the-wild evaluation: 100% accuracy was obtained on each of three in-the-wild face-swap deep fakes downloaded from YouTube.The videos involved Steve Buscemi/Jennifer Lawrence, Tom Cruise/Bill Hader, and Billie Eilish/Angela Martin identity combinations.
- Feature analysis: Behavior-Net similarities matched target behavior more strongly than source behavior, while comparisons across contexts showed identity-specific but somewhat context-dependent behaviors.Static FAb-Net similarities were nearly identical across same-person and different-person comparisons in the analyzed contexts, unlike Behavior-Net features.
- Reference-set analysis: With 2, 30, 50, 100, 1000, and 2000 WLDR clips, average detection accuracy was 65.4%, 92.2%, 93.2%, 94.0%, 97.3%, and 97.7%, respectively.Accuracy increased rapidly and then leveled off when the reference context was similar to the depicted context.
- Robustness: After recompression to qp = 40, average accuracy remained 94.5% on WLDR, 98.1% on FF, 93.2% on DFD, 80.9% on DFDC-P, and 93.3% on CDF.These results were described as almost identical to those for the higher-quality videos at the same threshold.
5 Discussion
The paper presents facial and behavioral biometrics as effective for detecting face-swap deep fakes and argues that the approach generalizes to unseen people and deep fakes. It may extend to puppet-master fakes but likely struggles with lip-sync manipulations.
- 5 Discussion: The technique combines facial and behavioral biometrics to detect face-swap deep fakes.It targets the mismatch between the depicted person and the identity or behavior the video purports to show.
- 5 Discussion: The approach generalizes well to previously unseen deep fakes and previously unseen people.
- 5 Discussion: The combined biometric should be able to detect puppet-master deep fakes because their underlying behavior does not match the depicted person.
- 5 Discussion: The method will likely struggle with lip-sync deep fakes, where only the mouth is modified to match new audio.The paper proposes learning inconsistencies between the mouth and the rest of the face or the underlying audio signal.
- 5 Discussion: Detection approaches may make convincing deep-fake creation more difficult and time consuming, although the threat will remain.