Shulou(Shulou.com)06/01 Report--

Today, I will talk to you about what FakeSpotter in DeepFake is like. Many people may not know much about it. In order to make you understand better, the editor summarized the following content for you. I hope you can get something according to this article.

1. Preface

In recent years, a variety of GAN networks have achieved great success in image generation, and then the existing detectors are not enough to fully face the challenges of GAN networks. In this paper, we propose a model to distinguish true and false faces based on human neural behavior. we speculate that each layer of neuron activation function may extract more tiny features, and these features are very important for true and false face recognition.

two。 Method 2.1 Insight

Neuron covering technology is widely used in the internal behavior of traditional DNN. When the output value is greater than the threshold, the activated neuron will be used as another form of input, and the learning content will be saved layer by layer in the network.

However, some previous work has been done to detect antagonistic examples of key activated nerve layers.

Our work is inspired by layered activated neurons that capture subtle features of input and can be used to find differences between real and synthetic facial images.

2.2 simulate neuronal behavior

FakeSpotter detection framework

The above image is a FakeSpotter detection framework, which is different from the traditional framework in that face analysis is based on the activation characteristics of neurons in each layer.

In traditional DNN, whether a neuron in each layer is activated depends on whether its output is higher than the threshold Threshold.

We propose a strategy to establish the threshold, and the formula is as follows

The strategy of establishing threshold proposed in this paper

Above the fraction is the sum of the output values of each neuron.

| | N | represents the total number of neurons in the current layer |

| | T | represents the number of inputs in the current layer |

Finally, this threshold is used to determine whether the neuron is activated or not.

Insert a picture description here

The following figure is an algorithm to describe these two strategies.

Algorithm 13. The other implementation detail optimizer is SGD with a momentum of 0.9 and an initial learning rate of 0.0001. The loss function adopts binary cross-entropy loss binary cross-entropy. The CNN architecture uses Vgg-Face, replacing the backbone network with ResNet50, and with our MNC strategy. A five-layer fully connected network is designed as the final two-classification network. 4. Experimental performance

We evaluate the robustness of the model by compressing, blurring, scaling and adding noise.

The robustness of the model is evaluated by compressing, blurring, zooming and adding noise.

You can see that it doesn't perform so well on the dataset of testing DFDC.

Because there are two types of face replacement and sound replacement in this data set, and sound substitution is beyond the scope of FakeSpotter image-based detection framework.

Other detection models add the performance of the four operations mentioned above to the Cele-DFv2 dataset

It can be seen that FakeSpotter still maintains a good detection rate.

In the field of DeepFake detection, a key problem is the robustness of the model. A trained model may fail if it is changed to another data set. Inspired by the DNN neuron activation layer, this work averages the output value of the activation layer to each neuron and adds it as a threshold to the whole network for training. finally, several model experiments show that this threshold strategy based on the activation layer can obtain more subtle features and further improve the robustness of the model.

After reading the above, do you have any further understanding of the FakeSpotter in DeepFake? If you want to know more knowledge or related content, please follow the industry information channel, thank you for your support.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.