DeepFake-o-meter: An Open Platform for DeepFake Detection (2024)

I Introduction

The buzzword DeepFakes has been frequently featured in the news and social media to refer to realistic impersonating images, videos, and audios that are generated using AI algorithms. Although fabrication and manipulation of digital media are not a new phenomenon [1], powerful AI technology, in particular, deep neural networks (DNNs), and the unprecedented computing power have made it easier than ever to create sophisticated and compelling fakes.Left unchecked, DeepFakes can escalate the scale and danger of disinformation, and fundamentally erode our trust in digital media.

In this work, we describe the design and the underlying mechanism of DeepFake-o-meter in details. We start with the overall architecture of the system, which is composed of a web-based front-end that interacts with the user, and an on-server back-end to perform analyses on the input videos. The separation of front-end and back-end is to ensure the security of the user-uploaded data, as well as to accommodate the long running time and the short response time to the users. We further provide an overview of the DeepFake detection algorithms that have been integrated into the current DeepFake-o-meter system. All these algorithms are recent and represent the state-of-the-art in DeepFake detection (two of the algorithms are from the top-performers of the Global DeepFake Detection Challenge). DeepFake-o-meter is designed to be an open-architecture, which can be augmented by incorporating more detection methods over time. We describe the API structures that are needed for third-party developers of DeepFake detection algorithms to have their method integrated into DeepFake-o-meter.

II Platform Design

This section describes the architecture design of deepfake-o-meter platform. Our platform is composed by three components: Front-end, Back-end and Data synchronizing. The front-end is the website portal to interact with users. The back-end is the core component of this platform, which calls corresponding detection methods to analyze the submitted videos and the data synchronizing is the protocol for exchanging the interested data between front-end and back-end. The overview of the platform architecture is illustrated in Fig.1.

II-B Back-end

The back-end is a computation server mainly for performing deepfake detection methods. In this section, we will describe the key design, returned results and integrated deepfake detection methods respectively.

II-B1 Key design

Once the user submits video from the front-end, the back-end starts to call corresponding detection methods for the submitted video. However, different detection methods depend on different environment settings and different detection methods are designed using different programming styles. Therefore, we design an unified framework to integrate the mainstream deepfake detection methods. Specifically, our framework has two major designs, Container and Coding structure to handle the diversity of environment and programming styles respectively.

Container.We know virtual machines are the first generation tools to solve the environmental conflict problem on a single machine. However, due to the heavy resources occupation, redundant operation and slow startup, virtual machines are replaced by Containers, which can isolate the process without creating a simulate operating system.Docker²²2https://www.docker.com/ is the most popular container solution currently, allowing developers to package their applications and dependent environment into a portable container, which can be run on any other machines.To freely run each method, we independently create docker image for each detection method.

Coding structureIn order to maintain each method efficiently, we design a coding structure that can provide an interface for each method to follow. Specifically, we design a base class containing four basic functions, named run, crop_face, preproc, postproc, get_softlabel and get_hardlabel.

1. -

   run: This is the entrance function to process an input image. The input argument is an image and output is the detection score. Given the input image, this function will internally call the functions crop_face, preproc, postproc, get_softlabel, get_hardlabel in sequel.

2. -

   crop_face: Since many methods require to extract the face area from the input image before prediction, this function provides an interface to wrap up the face extraction process. This function is optional.

3. -

   preproc: After face extraction, many methods apply pre-processing operations to the input face, such as changing the channel order or color space. Therefore, the pre-processing operations can be put in here. This function is also optional.

4. -

   get_softlabel: This function takes as input the prepossessed face and outputs the confidence score (soft label). Less score denotes the face is faker. The details of calling specific detection methods are wrapped here.

5. -

   get_hardlabel: Based on the soft label, this function assigns the input to real or fake label.

See Also

Deepfakes, explained | MIT Sloan

The code structure will also be exposed to the researchers who would like to integrate new methods into platform. The researchers can follow the structure to split their codes into different functions accordingly.

II-B2 Returned results

The formatting of returned results is also an important point. For better visualization to users, we curve the score of each face along with the corresponding frame and save the prediction of each frame to a video. Besides the visualization, we also sort the score along all frames and calculate the Area Under Curve (AUC) score. The results will be zipped together and sent back to the front-end for user to download. Fig.3 illustrates several examples of the returned results. The left part is submitted video and right part plots the corresponding score. Note our platform supports to run several methods at the same time, thus the bottom two examples contains multiple curves.

II-B3 DeepFake detection methods

Our platform integrates the following deepfake detection methods into this platform.

1. 1.

   MesoNet [2] is a self-designed CNN model that focuses on the mesoscopic properties of images. They provide two variants of MesoNet, namely, Meso4 and MesoInception4. Meso4 uses conventional convolutional layers, while MesoInception4 is based on the more sophisticated Inception modules [3]. We integrate MesoInception4 into the platform.

2. 2.

   FWA [4] is based on ResNet-50 [5] which detects DeepFake videos by exposing the face warping artifacts due to the resizing and interpolation operations.

3. 3.

   VA [6] targets the visual artifacts in the face organs such as eyes, teeth and facial contours of the synthesized faces. Two variants of this method are provided: VA-MLP and VA-LogReg. VA-MLP is based on a crafted CNN, and VA-LogReg uses a simpler logistic regression model. We integrate VA-MLP into the platform

4. 4.

   Xception [7] comes with FaceForensics++ dataset. It corresponds to a DeepFake detection method based on the XceptionNet model [8]. This method provides three variants: Xception-raw, Xception-c23 and Xception-c40. Xception-raw are trained on raw videos, while Xception-c23 and Xception-c40 are trained on compressed videos with different degrees, respectively. We integrate Xception-c23 into the platform.

5. 5.

   ClassNSeg [9] is another CNN based DeepFake detection method that is formulated to a multi-task learning problem to imultaneously detect forgery images and segment manipulated areas.

6. 6.

   Capsule [10] employs the VGG19 [11] capsule structure [12] as the backbone architecture for DeepFake classification.

7. 7.

   DSP-FWA is a further improved method based on FWA, which incorporates a spatial pyramid pooling (SPP) module [13] to better tackle the variations of face resolutions.

8. 8.

   CNNDetection [14] utilizes a standard image classifier trained on only ProGAN [15] , finding it generalizes surprisingly well to unseen architectures, datasets, and training methods.

9. 9.

   Upconv [16] argues that common up-sampling methods (upconvolution or transposed convolution) lack the ability to reproduce spectral distributions of natural training data correctly. They take $2D$ amplitude spectrum as feature and utilize a basic SVM classifier.

10. 10.

    WM ensembles two WS-DAN [17] models (with EfficientNet-b3 [18] and Xception [8] feature extractors, respectively) and a Xception classifier to produce per-face predictions.

11. 11.

    Selim utilizes state-of-the-art encoder, EfficientNet B7, pretrained with ImageNet [19] and noisy student [20], and uses a heuristic way to select 32 frames for each video to average predictions.

The summary of each detection method with code repositories is given in Table I.

Methods	Repositories	Release Date
MesoNet [2]	https://github.com/DariusAf/MesoNet	2018.09
FWA [4]	https://github.com/danmohaha/CVPRW2019_Face_Artifacts	2018.11
VA [6]	https://github.com/FalkoMatern/Exploiting-Visual-Artifacts	2019.01
Xception [7]	https://github.com/ondyari/FaceForensics	2019.01
ClassNSeg [9]	https://github.com/nii-yamagishilab/ClassNSeg	2019.06
Capsule [10]	https://github.com/nii-yamagishilab/Capsule-Forensics-v2	2019.10
CNNDetection	https://github.com/peterwang512/CNNDetection	2019.12
DSP-FWA	https://github.com/danmohaha/DSP-FWA	2019.11
Upconv	https://github.com/cc-hpc-itwm/UpConv	2020.03
WM	https://github.com/cuihaoleo/kaggle-dfdc	2020.07
Selim	https://github.com/selimsef/dfdc_deepfake_challenge	2020.07

III Conclusion

In this work, we describe an open platform, known as DeepFake-o-meter, for DeepFake detection. This platform is composed by front-end and back-end. The front-end is a web application to interact with users and back-end perform corresponding detection methods on submitted videos. The platform integrates more than 10 state-of-the-art detection methods and it also provides interfaces for researchers to incorporate their method into the platform.

For future works, we will continue integrate more DeepFake detection methods into the platform. Furthermore, we will study the use of multi-GPU platform to accelerate the analysis process. We will also augment the APIs so as to accommodate more general detection methods for other media formats (still images and audio signals).

Acknowledgment. This work is partly supported by the research project of National Science Foundation (no. IIS-2008532).

References

• [1]H.Farid, Digital Image Forensics.MIT Press, 2012.
• [2]D.Afchar, V.Nozick, J.Yamagishi, and I.Echizen, “Mesonet: a compact facialvideo forgery detection network,” in IEEE International Workshop onInformation Forensics and Security (WIFS), 2018.
• [3]C.Szegedy, W.Liu, Y.Jia, P.Sermanet, S.Reed, D.Anguelov, D.Erhan,V.Vanhoucke, and A.Rabinovich, “Going deeper with convolutions,” inCVPR, 2015.
• [4]Y.Li and S.Lyu, “Exposing deepfake videos by detecting face warpingartifacts,” in IEEE Conference on Computer Vision and PatternRecognition Workshops (CVPRW), 2019.
• [5]K.He, X.Zhang, S.Ren, and J.Sun, “Deep residual learning for imagerecognition,” in CVPR, 2016.
• [6]F.Matern, C.Riess, and M.Stamminger, “Exploiting visual artifacts to exposedeepfakes and face manipulations,” in IEEE Winter Applications ofComputer Vision Workshops (WACVW), 2019.
• [7]A.Rössler, D.Cozzolino, L.Verdoliva, C.Riess, J.Thies, andM.Nießner, “FaceForensics++: Learning to detect manipulated facialimages,” in ICCV, 2019.
• [8]F.Chollet, “Xception: Deep learning with depthwise separable convolutions,”in CVPR, 2017.
• [9]H.H. Nguyen, F.Fang, J.Yamagishi, and I.Echizen, “Multi-task learning fordetecting and segmenting manipulated facial images and videos,” inIEEE International Conference on Biometrics: Theory, Applications andSystems (BTAS), 2019.
• [10]H.H. Nguyen, J.Yamagishi, and I.Echizen, “Use of a capsule network todetect fake images and videos,” arXiv preprint arXiv:1910.12467,2019.
• [11]K.Simonyan and A.Zisserman, “Very deep convolutional networks forlarge-scale image recognition,” arXiv preprint arXiv:1409.1556, 2014.
• [12]S.Sabour, N.Frosst, and G.E. Hinton, “Dynamic routing between capsules,”in NeurIPS, 2017.
• [13]K.He, X.Zhang, S.Ren, and J.Sun, “Spatial pyramid pooling in deepconvolutional networks for visual recognition,” IEEE transactions onpattern analysis and machine intelligence (TPAMI), 2015.
• [14]S.-Y. Wang, O.Wang, R.Zhang, A.Owens, and A.A. Efros, “Cnn-generatedimages are surprisingly easy to spot… for now,” in CVPR, 2020.
• [15]T.Karras, T.Aila, S.Laine, and J.Lehtinen, “Progressive growing of GANsfor improved quality, stability, and variation,” in ICLR, 2018.
• [16]R.Durall, M.Keuper, and J.Keuper, “Watch your up-convolution: Cnn basedgenerative deep neural networks are failing to reproduce spectraldistributions,” in CVPR, 2020.
• [17]T.Hu, H.Qi, Q.Huang, and Y.Lu, “See better before looking closer: Weaklysupervised data augmentation network for fine-grained visualclassification,” arXiv preprint arXiv:1901.09891, 2019.
• [18]M.Tan and Q.Le, “Efficientnet: Rethinking model scaling for convolutionalneural networks,” in ICML, 2019.
• [19]J.Deng, W.Dong, R.Socher, L.-J. Li, K.Li, and L.Fei-Fei, “ImageNet: ALarge-Scale Hierarchical Image Database,” in CVPR, 2009.
• [20]Q.Xie, M.-T. Luong, E.Hovy, and Q.V. Le, “Self-training with noisy studentimproves imagenet classification,” in CVPR, 2020.