Source-linked AI summary
WildDeepfake: A Challenging Real-World Dataset for Deepfake Detection
Bojia Zi, Minghao Chang, Jingjing Chen, Xingjun Ma, Yu-Gang Jiang
TL;DR
Existing deepfake datasets often use limited scenes, actors, and researcher-generated forgeries, motivating better evaluation on internet deepfakes. The paper introduces WildDeepfake, evaluates baseline and attention-based detectors, and finds substantially harder detection on WildDeepfake while ADDNets improve performance. However, ADDNet-2D reaches only 76.25% accuracy on WildDeepfake.
Problem
Existing datasets use limited real-video settings and researcher-crafted virtual deepfakes, leaving real-world internet deepfakes insufficiently represented for detector development and evaluation.
Method
The paper collects WildDeepfake from the internet, systematically evaluates baseline detectors, and proposes 2D and 3D ADDNets that use attention masks to adjust facial features.
Results
Baseline accuracy decreases drastically on WildDeepfake, with no baseline above 70%, while ADDNet-2D outperforms the best baseline by 7% and ADDNet-3D reaches 65.50%.
Takeaways & Limitations
WildDeepfake provides a challenging real-world supplement for developing and evaluating deepfake detectors, and attention-based feature adjustment improves detection across the evaluated datasets.
Takeaways & Limitations
ADDNet-2D does not fully solve wild-deepfake detection, achieving only 76.25% accuracy on WildDeepfake.
Abstract
from arXiv · showhide
In recent years, the abuse of a face swap technique called deepfake has raised enormous public concerns. So far, a large number of deepfake videos (known as "deepfakes") have been crafted and uploaded to the internet, calling for effective countermeasures. One promising countermeasure against deepfakes is deepfake detection. Several deepfake datasets have been released to support the training and testing of deepfake detectors, such as DeepfakeDetection and FaceForensics++. While this has greatly advanced deepfake detection, most of the real videos in these datasets are filmed with a few volunteer actors in limited scenes, and the fake videos are crafted by researchers using a few popular deepfake softwares. Detectors developed on these datasets may become less effective against real-world deepfakes on the internet. To better support detection against real-world deepfakes, in this paper, we introduce a new dataset WildDeepfake which consists of 7,314 face sequences extracted from 707 deepfake videos collected completely from the internet. WildDeepfake is a small dataset that can be used, in addition to existing datasets, to develop and test the effectiveness of deepfake detectors against real-world deepfakes. We conduct a systematic evaluation of a set of baseline detection networks on both existing and our WildDeepfake datasets, and show that WildDeepfake is indeed a more challenging dataset, where the detection performance can decrease drastically. We also propose two (eg. 2D and 3D) Attention-based Deepfake Detection Networks (ADDNets) to leverage the attention masks on real/fake faces for improved detection. We empirically verify the effectiveness of ADDNets on both existing datasets and WildDeepfake. The dataset is available at: https://github.com/OpenTAI/wild-deepfake.
1 INTRODUCTION
Deepfake detection needs real-world evidence because existing datasets rely on limited actors, scenes, and researcher-generated forgeries. WildDeepfake addresses this gap with an internet-collected benchmark, systematic evaluations, and attention-based detectors.
- Deepfakes are deep learning-based facial forgeries that swap one person’s face with another in video, creating risks including political distress, blackmail, and fake terrorism events.
- Existing datasets commonly use a few volunteer actors in limited scenes and researcher-crafted virtual deepfakes made with several popular software tools.
- Virtual deepfakes lack the diversity of wild videos, which can contain many people, varied scenes, and forgeries produced by unknown software types, versions, or combinations.
- These dataset weaknesses may prevent detectors trained on virtual deepfakes from fully generalizing to real-world deepfakes.
- WildDeepfake is collected completely from the internet, while baseline networks and 2D/3D Attention-based Deepfake Detection Networks are evaluated on existing and WildDeepfake datasets.
- The study reports that baselines perform well on existing datasets but poorly on WildDeepfake, whereas ADDNets are empirically effective across both dataset types.
2 RELATED WORK
Related work covers deepfake generation and detection methods, including face-swap pipelines that use learned representations and attention masks. The paper positions its evaluation alongside established CNN, XceptionNet, MesoNet, and sequence-based approaches.
- Deepfake generation commonly uses GANs and open-source software such as Faceswap-GAN and Faceswap.
- Encoder-decoder face-swap systems learn shared source and target features, generate a swapped face, and use an attention mask during fusion to improve realism.
- MesoNet examines mesoscopic face-image properties, while CNN-LSTM methods use frame features and sequence prediction to detect deepfake videos.
- The study evaluates XceptionNet, MesoNets, and conventional CNN networks on existing datasets and WildDeepfake.
3 DATASETS FOR DEEPFAKE DETECTION
Existing deepfake datasets are largely built by manipulating collected source videos, whereas WildDeepfake collects real and fake videos from the internet. Its diversity, realism, and labor-intensive annotation target real-world detection challenges.
- Existing Deepfake Datasets: Existing datasets such as UADFV, Deepfake-TIMIT, and FaceForensics++ generate deepfake videos from real videos using specified face-manipulation techniques.
- Dataset Motivation: WildDeepfake is intended to supplement existing datasets because virtual-dataset flaws may limit detector generalization to wild deepfakes.
- WildDeepfake Dataset: WildDeepfake begins with over 1,200 internet videos, removes traditional manipulations and videos without identifiable real versions, and retains 707 deepfake videos.
- WildDeepfake Dataset: WildDeepfake contains 1,180,099 face images across 7,314 annotated face sequences from 707 videos, after discarding unknown and disagreement-labeled sequences.
- WildDeepfake Dataset: The 7,314 sequences are split into 6,508 training and 806 testing sequences based on sequence similarities to separate training and test faces to some extent.
- WildDeepfake Dataset: The dataset includes diverse activities, scenes, backgrounds, illumination, compression rates, resolutions, and formats, with well-made deepfake videos.
4 PROPOSED ADDNETS FOR DEEPFAKE DETECTION
The paper formulates deepfake detection as binary classification and introduces 2D and 3D attention-based networks that adjust features using facial attention masks. The 2D model handles individual faces, while the 3D model processes face sequences with shared attention blocks and convolutional classification.
- 4.1 Problem Formulation: Deepfake detection trains a CNN-based binary classifier to map extracted face inputs to real or fake labels by minimizing classification error.Real videos have class label 0 and deepfake videos class label 1; cross-entropy is an example loss function.
- 4.2.1 ADDNet-2D: ADDNet-2D takes a face image and its generated attention mask, applies an ADD block followed by a 2D CNN and classification layer.The network outputs probabilities for real (class 0) and fake (class 1), while intermediate features are adjusted using scaled attention masks.
- 4.2.2 ADDNet-3D: ADDNet-3D processes face sequences and corresponding attention-mask sequences through multiple shared-weight ADD blocks before 3D CNN classification.The block outputs are concatenated and reshaped, with one block processing each face–mask pair.
- 4.2.3 Attention Mask Generation Module: The attention mask generation module detects 68 facial landmarks, aligns the face, creates face and organ masks, and smooths their edges with Gaussian blur.The organ mask covers the eyes, nose, and mouth.
- 4.2 Proposed Detection Networks: Attention-based feature adjustment distinguishes ADDNets from existing detection networks by applying scaled attention masks at multiple ADD-block layers.The approach targets facial regions represented by attention masks while retaining the XceptionNet-based ADD-block structure.
- 4.2.3 Attention Mask Generation Module: Sequence-level detection can alternatively use ADDNet+LSTM by replacing the 3D CNN in ADDNet-3D with an LSTM network.Both designs retain the attention-based processing of face sequences and masks.
5 EXPERIMENTS
The experiments compare baseline and attention-based detectors across six datasets at image and sequence levels. Existing datasets yield strong baseline accuracy, but performance drops sharply on WildDeepfake; ADDNet-2D improves results there, while 3D methods remain less effective.
- Experimental Setup: The evaluation tests image- and sequence-level detectors on six datasets, including WildDeepfake and quality variants of DF-TIMIT and FF++.The datasets are DFD, DF-TIMIT LQ/HQ, FF++ LQ/HQ, and WildDeepfake.
- Image-level Detection: Existing datasets produce high baseline accuracies, including 99.91% on DF-TIMIT HQ and 99.62% on FF++ HQ.The best baseline reaches 90.02% on DFD, 99.65% on DF-TIMIT LQ, 90.91% on FF++ LQ, and 99.62% on FF++ HQ.
- Image-level Detection: No 2D baseline exceeds 70% accuracy on WildDeepfake, contrasting with their strong performance on existing datasets.XceptionNet is the best baseline on three of the five existing datasets.
- Image-level Detection: ADDNet-2D outperforms XceptionNet by 11% on DFD and 7% on WildDeepfake, while remaining within 1% of the best baseline on other existing datasets.Its WildDeepfake accuracy is 76.25%, although the challenge is not fully resolved.
- Sequence-level Detection: 3D baselines exceed 89% on the four DF-TIMIT and FF++ datasets but fall below 63% on WildDeepfake.I3D is strongest on most tested datasets, while C3D performs best on DFD; P3D performs worst among the baselines.
- Sequence-level Detection: ADDNet-3D reaches 65.50% on WildDeepfake, about 2% above I3D but below ADDNet-2D's 76.25%.The authors suggest frame-generation artifacts may distort temporal information inconsistently across frames or videos.
6 CONCLUSION
WildDeepfake is an internet-collected dataset designed to test deepfake detectors under more diverse real-world conditions. Baseline performance drops sharply on it, while ADDNet-2D performs comparably or better than state-of-the-art methods across the evaluated datasets.
- Dataset: WildDeepfake contains 1,180,099 images from 7,314 face sequences extracted from 707 deepfake and real videos.The videos were collected completely from the web and include diverse scenes, faces, activities, and high-quality deepfakes.
- Findings: WildDeepfake is more challenging than existing virtual datasets, with baseline detector performance decreasing drastically.The dataset is intended as a supplement for developing and evaluating detectors against real-world deepfakes.
- Proposed Method: ADDNet-2D achieves better or at least comparable performance to state-of-the-art methods across all evaluated datasets.The conclusion attributes this performance to attention-based feature adjustment for deepfake detection.