Deepfake is one of the most popular application of Deep Learning. But did you ever think why deepfake is getting popularity? Let me tell you why deepfake is getting the popularity.
Deepfake has the most magical application that can make a human fool. So, to learn about how deepfake works and what are the technology that work behind it let's get started with the new short video on How deepfakes work? So first we will see, what are deepfakes?
Why is gaining popularity in the market? Right. So deepfake is basically the phenomenon gained its name from the platform reddit, so reddit is one of the most popular platform who went by the name deepfakes but why it's called deepfakes right?
So, if we divide that name that we got that deep learning plus fakes right. So, these are the main two parts of calling this application as a deepfakes. So deep learning is the part of artificial intelligence, right?
Like where we actually makes our brains works. So, in this case, we actually, in the deep learning, we've used our brain. Right.
So, it's not the human brain. It's a machine brain. So, we use machine brain and make a brain like human and we use that to do some applications.
So, this is deepfakes. Is the application of deep learning. The first deepfake shared by that person were of unknowing celebrities with their faces attached to the bodies of Pornstars.
The first targets of deepfake were famous people including actors like Emma Watson, singers like Katy Perry and politicians president Obama. This is how deepfakes came into the picture. And these are the first very initial applications of deepfake.
Now we will have a look at. What makes deepfakes so special and why they feel that deepfakes is one of the most coolest applications. Right?
So as good as original and deepfake are very hard to distinguish from the original. If we don't see it with our own eyes, sometimes we believe it to exist or even to be true. So, from this, we actually can conclude that.
We can see that these deepfakes that trick people into believing something from the pictures, from the images you can see that's the difference between the origin and the generated one is very mixed. So, in this case, if you don't know the person. So, in that case, maybe you can end up believing that this person actually exists.
Now we will have a look at what's behind the deepfake. What is the algorithm we are using behind this deepfake? We already know deepfake is a part of deep learning but what algorithm we use over there.
So previously we used to do this, like changing the face or replacing some faces with bodies. What we used to do the Photoshop but the introduction of GAN changed the whole scenario. but what is GAN?
GAN is basically General Adversarial Network and we will see how GAN works. So, what is GAN, GAN is a generative adversarial network which has two parts one is Generator, one is Discriminator, right? So, what is Generative.
Describe how input can be generated from the input data set in terms of probabilistic model. So, GAN is based on probability now we will see the adversarial part train the network in the adversarial way. It's two of types.
We have two types. One is generator, one is discriminator now we will have a look at Network so Deep neural network are used for the training purpose So, this is the scenario when we call it scam. Why, what is the stance for ‘G’ ‘A’ ‘N’.
So, G is for Generative, A is Adversarial way of train your model and N is Network which is Deep Neural Network right in this, the note is the, instead of predicting a level, given certain features, they attempt to predict feature given a certain level, right? Now, we will have a look at, how GAN works? Gan uses unsupervised process to discover and learn automatically from input data to generate new data from the original data set.
So, you can see in the image, what do we do? We first give a noise to our generator. So, our Generator first gets a noise, but our generator does not have idea about what is real image right?
Now, our generator actually generated some samples from that. And what we give discriminator two parts, one is the generated image from generator one is the real image from that our discriminator used to say this is a real image and this is a fake image, right? So, this is how GAN works.
Now we will see like drill down to generator. So Generative model stands for probability and stands for joint probability, right? So, we all familiar with what is joint probability.
So joint probability means likely hood of two events will happen at the same time, two events should be independent to each other. That is outcome of event X does not influence the outcome of event Y. So, this is how our Generator of the GAN actually works.
And let me give you an example for that supposed to two dice are there and rolling. One of them is not depending on each other, right? So, this is that terms actually, we use in probability.
That is a joint and generator use that concept only Now, we will have a look at the generator objective pros and cons. So basically, the objective of generator is to model the data generation creates new data points from the given training data set and what is the flows we have in the generator. Model have the proper knowledge and data distributions.
Right? Cause the large amount of data is needed expensive to implement so if we go for any GAN application. We need to have a large amount of data, right?
and the process is actually, like expensive. So generative model learns the distribution of that data. It has a very good idea about what is your distribution of your training data set, right?
Now we will have a look at discriminator the second part of GAN. So Discriminative model in terms of probability, is a conditional probability basically GAN generator is a joint probability and discriminator is conditional probability So, in the conditional probability two events should be integrated into each other? That is outcome of X, does not influence the outcome of Y conditional probability of B the probability of the occurred event A Given that B has already happened.
So, this is the concept behind the discriminator, right? Now we'll have a look at the objectives pros and cons of the discriminator. So, what is the objective of the discriminator?
The discrimnator tries to learn using probabalistic approaches Pros, to train a model, is comparetively easy. Cons, Tries to classify the data from the training set. So, discriminated modern learns the boundaries of the tasks.
So basically, if we take a look at the real scenario, so the discriminator help us to understand what image is actually a real image and what is actually a fake image, right? So in this case, we used generator for that and with time our generator and discriminator becomes very efficient for that. Our generator is able to make a data image like an actual image, right?
and our discriminator is also become very. Efficient enough with time that discriminator can say image you have made, this is a fake image the image I’m getting from the actual image this is the true image, right? So this is how our GAN work and with time the way we train our GAN like with generator and discriminator are getting trained in the same time, right?
So these are the key points off GAN. So now we will have a look at to the steps used in GAN, so the generator takes the random numbers, yes, we put a noise to our generator and it returns an image the image from that. It returns an image that is a noise image, right?
Though, we get from the generator when we initially start our process. The generator image is then fed to discriminator alongside of batch of images taken from the actual ground truth data set. So in the first case, we'll give this discriminator two inputs that is one is generated image from generator and second one is the from the actual real image.
The discriminator takes both real and fake image and returns probabilities With one or zero one is basically representing the prediction of Image being authentic and zero represents the image is fake. So in GAN, we have a double feedback loop, right? And the discriminator of GAN is in the feedback loop with ground truth of images which can already know.
And the generator of GAN is in a feedback loop with the discriminator. So we used to loops one is for generating and one is for discriminator. So this is how actually our deepfakes works.
So if you want to have more advanced topics like this, then please subscribe to the Great learning channel. So if you want to learn some advanced topic like this, then please go to Great Learning Academy Platform. Where you will get almost 80 plus free courses and after completing your course you can claim your certificate as well.
And if you want to do this course on your mobile application that is also possible using Great Learning app. Thank you so much.