Source-linked AI summary
Evaluation of an Audio-Video Multimodal Deepfake Dataset using Unimodal and Multimodal Detectors
Hasam Khalid, Minha Kim, Shahroz Tariq, Simon S. Woo
TL;DR
Deepfake detection research and datasets have largely focused on either video or audio, despite the challenge of deepfakes combining both. The paper evaluates unimodal, ensemble, and multimodal methods on FakeAVCeleb; ensembles perform best, while multimodal methods perform worst. The authors also identify the absence of hyperparameter tuning for their baselines as a scope limitation.
Problem
Existing deepfake datasets and detectors largely address video or audio separately, and DFDC does not label which modality is fake.
Source: §1 · p. 2Method
The authors evaluate unimodal, ensemble-based, and multimodal detection baselines on FakeAVCeleb, which contains deepfake video with synthesized audio.
Source: §1 · p. 2Results
Ensemble-based methods outperform unimodal methods, while multimodal methods perform worst; Xception is most stable (Audio: 76% and Video: 73%).
Source: §4.3.3 · p. 6, §5 · p. 7Takeaways & Limitations
The authors conclude that multimodal deepfake detection is not trivial and requires further research.
Source: §1 · p. 2The baseline methods were not hyperparameter-tuned, which the authors leave outside the paper's scope.
Source: §4.3.1 · p. 6
Abstract
from arXiv · showhide
Significant advancements made in the generation of deepfakes have caused security and privacy issues. Attackers can easily impersonate a person's identity in an image by replacing his face with the target person's face. Moreover, a new domain of cloning human voices using deep-learning technologies is also emerging. Now, an attacker can generate realistic cloned voices of humans using only a few seconds of audio of the target person. With the emerging threat of potential harm deepfakes can cause, researchers have proposed deepfake detection methods. However, they only focus on detecting a single modality, i.e., either video or audio. On the other hand, to develop a good deepfake detector that can cope with the recent advancements in deepfake generation, we need to have a detector that can detect deepfakes of multiple modalities, i.e., videos and audios. To build such a detector, we need a dataset that contains video and respective audio deepfakes. We were able to find a most recent deepfake dataset, Audio-Video Multimodal Deepfake Detection Dataset (FakeAVCeleb), that contains not only deepfake videos but synthesized fake audios as well. We used this multimodal deepfake dataset and performed detailed baseline experiments using state-of-the-art unimodal, ensemble-based, and multimodal detection methods to evaluate it. We conclude through detailed experimentation that unimodals, addressing only a single modality, video or audio, do not perform well compared to ensemble-based methods. Whereas purely multimodal-based baselines provide the worst performance.
1 INTRODUCTION
Deepfake generation across images, videos, and audio raises security and privacy concerns, while existing datasets and detection methods largely address only one modality. The paper evaluates unimodal, ensemble, and multimodal baselines on FakeAVCeleb and reports that ensembles perform best, whereas unimodal and multimodal approaches have substantial difficulty.
1 INTRODUCTION
- Deepfake generation has expanded from face manipulation to realistic cloned voices, increasing the potential for impersonation and security and privacy harms.The paper notes risks including forged videos used to portray false messages and spread misleading information.Source: §1 · pp. 1–2
- Existing video datasets lacked corresponding fake audio, while audio datasets lacked corresponding lip-synced or deepfake video; DFDC also labeled an entire pair fake when either modality was fake.That labeling did not identify whether the audio, video, or both were fake.Source: §1 · p. 2
- The authors evaluate FakeAVCeleb, which contains deepfake videos with lip-synced synthesized audio, using unimodal, ensemble, and multimodal baselines.The tested baselines include Meso-4, MesoInception-4, Xception, EfficientNet-B0, VGG16, and three open-source multimodal methods.Source: §1 · p. 2
- Ensemble-based methods performed best on audio-video deepfakes, but their prediction accuracy remained below 85%.The authors conclude that multimodal deepfake detection is not trivial and requires further research.Source: §1 · p. 2
2 RELATED WORKS
Prior work spans video and audio datasets and detectors, but the reviewed datasets and multimodal methods leave gaps for evaluating deepfakes across both modalities. The related work describes the paper's focus on simultaneous video and fake-audio detection.
2.1.1 Deepfake Video Datasets.
- Video datasets progressed from small, low-quality UADF and Deepfake TIMIT collections to larger datasets such as FF++ and DFDC, with later datasets using newer generation methods.The review notes DFDC's environmental conditions can differ from real-world scenarios and make detection difficult.Source: §2.1.1 · pp. 2–3
2.1 Fake Media Datasets
- FakeAVCeleb is presented as the reviewed dataset containing real and fake videos with respective audios, unlike other newer datasets focused on video with real or no background audio.DFDC contains fake audio, but its labels do not specify whether the audio or video is fake, and its videos are not lip-synced with the audios.Source: §2.1.1 · p. 3
2.1.2 Fake Audio Dataset.
- The reviewed fake-audio resources include ASVspoof and Fake-or-Real, while FakeAVCeleb uses targeted voice cloning and Wav2Lip to produce lip-synced deepfakes.Fake-or-Real synthesizes a single man or woman's voice rather than cloning a specific person.Source: §2.1.2 · p. 3
2.2.1 Deepfake Video Detection.
- Video-detection research includes methods targeting splice inconsistencies, eye blinking, head poses, and data-driven traces, as well as transfer-learning and two-stream approaches.The cited examples include splice detection, abnormal eye-blinking, irregular head poses, transfer learning [11], and a two-stream network.Source: §2.2.1 · p. 3
2.2.2 Fake Audio Detection.
- Fake-audio detection research includes handcrafted cepstral features with GMM classifiers and deep-learning models using CNNs, RNNs, or their combination.FFT and MFCC are among the feature-extraction approaches described for deep-learning models.Source: §2.2.2 · p. 3
2.2.3 Multi-modal Detection.
- Multimodal detection includes methods based on audio-visual emotion or cross-modal dissimilarity, but the reviewed methods use datasets with deficiencies identified by the authors.The paper states that simultaneous multimodal detection of deepfake video and fake audio had not been explored, to the authors' knowledge.Source: §2.2.3 · pp. 3–4, Figure 2 · p. 4
3 MULTIMODAL DEEPFAKE DETECTION
The paper evaluates FakeAVCeleb with frame-based video detection and audio-video baselines organized as unimodal, ensemble, and multimodal methods. Its dataset includes celebrity videos across five ethnic backgrounds, with face-swapping, voice cloning, and facial reenactment used in its generation.
3.1 Dataset
- FakeAVCeleb's real videos come from VoxCeleb2 and include five ethnic backgrounds, with 600 unique videos averaging seven seconds.Each ethnic group has 100 real videos of 100 celebrities, with 50 of each gender; the dataset was not public, so the authors requested access.Source: §3.1 · p. 4
- The dataset combines face-swapped or reenacted videos with cloned audio, using Faceswap, FSGAN, DeepFaceLab, RTVC, and Wav2Lip.Wav2Lip reenacts the video based on source audio.Source: §3.1 · p. 4
3.2 Evaluation Methods
- The evaluation classifies extracted video frames as real or fake and compares unimodal, ensemble, and multimodal methods.The three categories are illustrated as one-modality models, separately trained models combined as an ensemble, and one model trained on both modalities.Source: §3.2 · p. 4
3.2.1 Unimodal Methods.
- Unimodal baselines evaluate audio or video separately and include VGG16, Meso-4, MesoInception-4, Xception, and EfficientNet-B0.The paper describes Meso-4 as focusing on mesoscopic features and MesoInception-4 as replacing its first convolutional layers with InceptionNet modules.Source: §3.2.1 · p. 4
- A unimodal classifier cannot consider both modalities, so when audio and video have different labels it can predict only from its own modality.For example, an audio-trained classifier predicts fake for fake audio with real video, whereas a video-trained classifier predicts real.Source: §3.2.1 · pp. 4–5
3.2.2 Ensemble Methods.
- Ensembles combine separately trained audio and video classifiers using soft voting, which averages class probabilities, or hard voting, which uses majority votes.The ensemble uses baseline models from the unimodal methods.Source: §3.2.2 · p. 5
3.2.3 Multimodal Methods.
- The evaluated multimodal methods include feature-attention and concatenation architectures, plus CDCN, which fuses low-, mid-, and high-level features for facial-depth prediction.The first two methods were originally applied to food-recipe classification and movie-genre prediction, respectively; CDCN was used for face anti-spoofing.Source: §3.2.3 · p. 5
4.1 Preprocessing
The authors preprocess video and audio separately, extracting frames and converting audio into MFCC feature images. They evaluate on held-out individuals using a balanced test set of real and fake videos.
4.1 Preprocessing
- Audio is sampled at 16 kHz and converted to 80-feature MFCC frames using a 25 ms Hann window, 10 ms shifts, and a 512-point FFT.The resulting MFCC features are stored as three-channel images for model input.Source: §4.1 · p. 5
- The general test set contains 70 real and 70 fake videos from individuals absent from training, yielding 1,895 real and 2,418 fake video frames.Audio preprocessing yielded 70 real and 70 fake MFCC feature images.Source: §4.1 · p. 5
4.2 Experiment Settings
The experiments compare unimodal, ensemble, and multimodal baselines on FakeAVCeleb, with separate unimodal training and paired audio-video ensemble and multimodal setups.
4.2 Experiment Settings
- The study evaluates unimodal, ensemble, and multimodal baselines using precision, recall, F1-score, and accuracy.Training uses Adam at a learning rate of 1 × 10−5, with early stopping after 10 epochs without improvement.Source: §4.2 · p. 5
4.2.1 Unimodal.
- Five classifiers are trained on 133,724 video frames or 720 MFCC images from 360 fake and 360 real videos.The video-only and audio-only settings train classifiers on their respective modalities.Source: §4.2.1 · p. 5
4.2.2 Ensemble.
- Ensembles pair one audio classifier with one video classifier and combine predictions using hard- or soft-voting.One MFCC image per video is upsampled to pair with 25 sampled frames; other imbalance-handling methods are outside the paper’s scope.Source: §4.2.2 · p. 5
4.2.3 Multimodal Audio+Video.
- Three existing multimodal methods are adapted to classify FakeAVCeleb’s video-frame and audio-MFCC modalities.The selected methods were originally developed for food-recipe classification, movie-genre prediction, and face anti-spoofing.Source: §4.2.3 · pp. 5–6
4.3 Unimodal Results
Unimodal baselines have difficulty distinguishing real and fake FakeAVCeleb media. Video and audio results vary across classifiers, and Xception is the most stable performer across modalities.
4.3 Unimodal Results
- Best unimodal detection reaches around 76% for audio and 81% for video, while results indicate difficulty distinguishing real from fake media.The reported test set contains real videos and all three types of fakes from FakeAVCeleb.Source: §4.3 · p. 6, §4.3.3 · p. 7
4.3.1 Results for Video-only Trained Classifier.
- 81% accuracy for VGG and 43% for Meso-4 mark the best and worst video-only results, respectively.The authors suggest Xception might outperform VGG with proper hyperparameter tuning, which they did not conduct.Source: §4.3.1 · p. 6
4.3.2 Results for Audio-only Trained Classifier.
- 76% for Xception and 50% for EfficientNet-B0 are the best and worst audio-only results, respectively.The authors note that none of the baselines provides satisfactory audio detection performance.Source: §4.3.2 · p. 6
4.3.3 Summary.
- Xception is the most stable performer, with 76% accuracy on audio and 73% on video.The authors characterize distinguishing real and fake FakeAVCeleb audio and video as relatively difficult for state-of-the-art deepfake detectors.Source: §4.3.3 · pp. 6–7
4.4 Ensemble Results
Among the tested audio-video ensembles, VGG performs best and Xception worst; soft- and hard-voting scores are reported as the same, and no ensemble exceeds 90%.
4.4 Ensemble Results
- 78% for the VGG ensemble is the best result, while the Xception ensemble is worst at 43%.MesoInception-4 is second best at 72%.Source: §4.4 · p. 7
- Soft- and hard-voting show the same prediction score, and none of the ensembles exceeds 90%.The authors attribute the lack of a voting-method difference to having only two classifiers in each ensemble.Source: §4.4 · p. 7
4.5 Multimodal Results
On FakeAVCeleb, Multimodal-2 achieved the highest reported accuracy among the three multimodal baselines, while Multimodal-1 and CDCN performed less well. The authors note that these methods may not be suited to deepfake detection and call for further research.
4.5 Multimodal Results
- 67.3% accuracy for Multimodal-2 exceeded Multimodal-1's 50% and CDCN's around 51% on FakeAVCeleb.All three models were trained for 50 epochs; Table 3 reports precision, recall, F1-score, and accuracy.Source: §4.5 · p. 7
- Multimodal-1 and CDCN performed less well than Multimodal-2, which the authors suggest may reflect their design for food-recipe classification and movie-genre prediction.The passage presents this as one possible reason for the performance difference.Source: §4.5 · p. 7
- The authors identify detecting deepfakes with only one of the two modalities fake as a challenge for multimodal methods and call for further research.Source: §4.5 · p. 7
5 CONCLUSION
Deepfakes that pair realistic video with lip-synced cloned voices make two-modality detection a challenging task. Experiments on FakeAVCeleb found that unimodal methods performed worse than ensemble-based methods, while multimodal methods performed worst.
5 CONCLUSION
- Deepfakes combining realistic video with perfect lip-synced cloned human voices pose a challenging detection task across two modalities.Source: §5 · p. 7
- Experiments on FakeAVCeleb found that unimodal methods performed worse than ensemble-based methods, while multimodal methods performed worst.Source: §5 · p. 7
- The paper reports experiments on the effects of multimodal deepfakes on existing state-of-the-art detection methods.Source: §5 · p. 7