Deepfakes are images and videos in which the original subject’s likeness is replaced with someone else’s using deep neural networks. There have been many recent advances in the development of deepfakes, the best deepfakes now are extremely realistic and undetectable by the human eye.
Deepfakes have the potential to severely disrupt the political, social, and economic sectors across the globe. They can be used to depict politicians and celebrities in compromising or pornographic images. In ordering to combat the threat that deepfakes pose, many governments and private companies are heavily investing in deepfake detection research. Google, Facebook, Microsoft, Amazon, and many universities have collaborated to create deepfake datasets and detection challenges to spur progress in this field.
This project will introduce a novel method for deepfake video detection utilizing dense optical flow and recurrent neural networks. The motivation behind this method is that the optical flows will characterize the motion of the video subject’s face and the recurrent neural network will learn about how this motion changes over time. Because deepfake generators work only to create a realistic image in the spatial domain for each frame individually, the recurrent neural network will be able to spot temporal inconsistencies or patterns in a deepfake video not found in real videos.
Below is a depiction of the overall pipeline used for deepfake detection:
The first part of this deepfake detector is the generation of optical flow fields. An optical flow field is a vector field showing the apparent motion between each frame of a video. Optical flow is calculated by assuming that the pixel colors associated with an object are persistent across each frame.
The optical flow fields were generated using the Farneback method [1] implemented in the OpenCV computer vision library. An example flow field showing the motion of cars on the highway is depicted below. The colors in the flow field are a function of the speed and direction of each object.
Dense optical flow fields were generated for the first 40 frames of each video. Each flow field was then decomposed into a magnitude and direction matrix. The matrices were then resized to have a shape of 200×200 and stacked on top of each other, creating a 200×200×2 array. Thus, for every video, a 40×200×200×2 array was generated. Stacking the matrices in this way makes sense because every stacked magnitude and direction matrix element is associated with the same pixel in the original video frame.
The second part of the detection system is a convolutional LSTM neural network [2]. The convolutional LSTM layers replace the matrix multiplication of a normal LSTM layer with convolution operations. Therefore the 200×200×2 magnitude and direction matrices can be inputted into the network without flattening them into a vector. The network structure is summarized below. The combined convolutional LSTM model can capture spatiotemporal correlations better than a sequential combination of a normal CNN and LSTM networks.
The neural network structure is composed of 3 convolutional LSTM layers, 3 batch normalization layers, 3 max pooling layers, a 512-neuron fully connected layer, and finally a sigmoid classification layer. The first 2 LSTM layers output a time series of arrays and the final LSTM layer outputs a single array which is flattened and fed into the fully connected layer before being classified. The full structure is depicted in Fig. 2. The Adam optimizer with a learning rate of 10-4 and binary cross-entropy loss function were used for training.
Optical flow fields were generated for 800 videos taken from the Celeb-DF deepfake database [3]. Half of these videos were real and the other half deepfakes. The model was trained on these videos using an 80/20 training/validation split. The model was then tested on the optical flows generated from 200 other videos also taken from Celeb-DF and equally split between deepfake and real videos.
| Model | Training acc. (%) | Validation acc. (%) | Testing acc. (%) |
|---|---|---|---|
| Conv-LSTM, 40 frames | 95.30 | 70.50 | 71.40 |
Below are plots of the model accuracy and loss:
Looking at the results it is clear that the model is overfitting to the training data based on the large discrepancy between training and validation accuracy. The loss plot illustrates how the model begins to overfit around epoch 11, this is when the validation loss starts increasing as the training loss rapidly falls.
In an attempt to improve the model accuracy a grid search was done over certain model hyperparameters (learning rate, batch size, hidden layer nodes) but no combination of parameters resulted in any improvement in accuracy.
After visualizing the generated optical flow fields, it became apparent that these optical flow fields were not consistent with what would be expected for subtle facial movements. The flow fields over the face contained a large amount of noise and were not smooth. This is likely a result of the inaccuracies associated with how the Farneback method tracks the pixels, as the facial pixels are all of similar color.
The Farneback method used to generate the optical flows was developed seventeen years ago and has been improved on by other methods. A new state-of-the-art method was employed to generate the flow fields, the LiteFlowNet convolutional neural network [4][5].
A face detector was also added using a Haar Cascade classifier [6] to extract only the optical flow over the facial region. This is to remove any noise associated with the background which is not useful for learning.
Above is a comparison of the optical flow generated between the same 2 frames using the Farneback method vs. LiteFlowNet visualized as vector fields, the red arrows represent the direction and magnitude of the change in position of each pixel from the previous frame. It's clear that the optical flow generated by LiteFlowNet is much smoother and less noisy than the Farneback method. The cleaner optical flow should make it easier for the neural network to learn to distinguish between real and deepfake videos.
Another issue is that when taking just one 40 frame clip from each video there were only 800 data points for training and validation, considering each clip as a data point. This is not enough data to train a deep neural network from scratch. In order to augment the dataset, all the videos were split into multiple clips. This increased the size of the data set by almost tenfold.
The next steps for this project are to generate the optical flows with the larger dataset using LiteFlowNet and added face extractor, and then train the convolutional LSTM network. Given a larger and better pre-processed dataset, the model should be able to better learn to distinguish between the real and fake videos and the accuracy should improve.
[1] G. Farnebäck, “Two-Frame Motion Estimation Based on Polynomial Expansion,” Image Analysis Lecture Notes in Computer Science, pp. 363–370, 2003.
[2] S. Xingjian, Z. Chen, H. Wang, D. Yeung, W. Wong, and W. Woo. “Convolutional LSTM network: A machine learning approach for precipitation nowcasting,” Neural Information Processing Systems, 2015.
[3] “Celeb-DF (v2): A New Dataset for DeepFake Forensics,” CVML Celeb-DF dataset, 16-Mar-2020. [Online]. Available: http://www.cs.albany.edu/~lsw/celeb-deepfakeforensics.html. [Accessed: 27-Mar-2020].
[4] T.-W. Hui, X. Tang, and C. C. Loy, “LiteFlowNet: A Lightweight Convolutional Neural Network for Optical Flow Estimation,” 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018.
[5] S. Niklaus, “A Reimplementation of LiteFlowNet Using PyTorch,” GitHub, 2019. [Online]. Available: https://github.com/sniklaus/pytorch-liteflownet. [Accessed: 23-Apr-2020].
[6] P. Viola and M. Jones, “Rapid object detection using a boosted cascade of simple features,” Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001.
[7] Knight, W., 2019. Facebook Is Making Its Own AI Deepfakes To Head Off A Disinformation Disaster. [online] MIT Technology Review. Available at: https://www.technologyreview.com/2019/09/05/65353/facebook-is-making-ai-deepfakes-to-head-off-a-disinformation-disaster/ [Accessed 24 November 2020].