Children Safety Retrieval (CENSER) System for Retrieval of Kidnapped Children from Brothels in India

Download Slides

1,628 kidnapped children were retrieved from a single railway station in northern India in 2018. These children were in the age group of 4 to 15 years and included 134 girls amongst which the youngest was only four years old. This has become an epidemic in India where gangs kidnap children from weaker sections of society, later to be sold either to brothels for prostitution or factories for child labor. The local law enforcement agencies only raid these places if they have confirmed identity for any of the children. Skylark labs ChildrEN SafEty Retrieval (CENSER) System is being used by non-profits in India to identify and then convince law enforcement to raid these places to retrieve the kidnapped children. The volunteers of the non-profits go to sell makeup to the brothels while wearing hidden cameras. The children buying makeup are recorded on the hidden cameras which are used by the CENSER system to establish a match with the database of missing children. Since the facial features of children have changed significantly since being kidnapped, the proposed system can perform age-invariant face recognition. A majority of the time, the image of the child is not available at the age when he/she was kidnapped. In that case, the CENSER system uses kinship analysis as well as matches with the sketch of the child. The talk will detail this very severe problem and how the CENSER system is making a change to save these kidnapped children.

Watch more Spark + AI sessions here
Try Databricks for free

Video Transcript

– Thank you very much for giving me the opportunity to present our work. Today I’ll be speaking about our CENSER, a platform which has been used by non-profits in India to retrieve children who have been kidnapped and pushed into brothels.

ChildrEN SafEty and Recovery (CENSER) System for abducted Children in india

And so I will speaking about that project.

So essentially the content of the talk would involve a brief introduction of what does the problem of human trafficking looks like in India and why it is such a severe problem. And then I’ll go into the child trafficking in a city called Varanasi, where we run most of our pilots, where the system is being used to retrieve these children. I’ll walk you through the brothels and the role of law enforcement in encouraging this process. And then I’ll also present some of the, statements of the victims who have been brutalized by these brothels. And then I’ll walk you through the non-profit, which is using our system to fight against this child trafficking and what are the primary challenge they were facing when they were trying to do it on their own, and how the system, which we have developed has aided them in challenging these brothels more effectively, and finally I’ll present the conclusions as to whether this project has been useful and how useful it has been for the, for the non-profits.

Human Trafficking in India

So human trafficking in India is a pandemic and it’s a pandemic all over the world, but especially in India, it’s a very big problem. And there are a few States in India, which you can see highlighted in black, which have been primarily responsible for most of the human trafficking cases. And when children are kidnapped from these weaker sections of India, they’re pushed into either forced labor or they are sexually exploited for prostitution. A lot of them they’re pushed into forced marriages, they are sold or they are used as servants in houses or forced into labor in different camps across India. So the place and that’s like a big problem. And as you can see the last year, I think it’s in 2016, almost 60,000 cases were reported of children being abducted and only a handful of them are ever retrieved and once they’re kidnapped.

So the place which you see on the map, which is marked with a star is a place called Varanasi, where a lot of the people from, a lot of the children from West Banglore, are pushed into the brokers over there. And we have primarily focused on that section because, one the NGO is there and second this is a very big profit area. So we wanted to see if the systems could be effectively used there to free the children.

Child Trafficking in Varanasi, India

So now I’ll work a bit through what these, what the child trafficking looks like in Varanasi, since we are focusing on that. So you see on the left is a typical image of a brothel in India, where you can go and you can see that are minor children who are standing outside and you can go and you can talk to their pimp and you can pay for their services. And the girls who are in this brothel, as I said, are primarily girls who are kidnapped from, parts of India, where people are very poor. And if you kidnap some of those children, they don’t really have the means or the resources or the know how to go and talk to law enforcement who can help their children. A lot of the time girls are lured into these brothels, forced by a promise of a job. And then they leave their home and then they are brought into these brothels and they’re forced and tortured to work over there. And also, I was very surprised to know that a lot of these girls are actually sold by their families to these brothels, because since they are in financial destitute, they get some money from these brothels for selling their children.

And finally, I always wondered as to everyone knows this is happening. These red light areas are in the center of the city, but there is no action from the locals or the law enforcement, because they are also involved in exploiting these minors. And that’s a major challenge as well in curbing this pandemic in India.

Through the Eyes of the Victims

So I can try to motivate this problem as much as I can, but I don’t think so, I can never do a good enough job myself. So I picked up some statements from the actual police reports, which were filed by some of the victims when they were retrieved from the brothels. And what are the circumstances they faced when they were in these brothels and how they were treated. So if you see this one statement on the top right, there is a girl who was 11 years old, was kidnapped from Uttar Pradesh a State in India. And she was pushed into one of these brothels, and then she was forced to do prostitution, she refused and, to make her comply, they burned her legs and the pierced the rods in her legs. And there’s another girl, was 15 years old, was again kidnapped and pushed into a brothel. And they tortured her, they beat her up, they busted open her head and they did not treat her well, which caused her worms in her head and after that, she was forced to do prostitution. So as you can see from these statements, the conditions in which these girls are kept, and the way they are treated are horrendous and they go through enormous torture in these brothels. So that was also a lot of our motivation to really work on this project and see that it gets to a sustainable point.

So and also when I visited these brothels and I looked at, the situation and I spoke to the girls who were retrieved, a natural thing came into my mind is that why doesn’t the law enforcement do anything? And the law enforcement doesn’t do anything primarily because it’s a lucrative trade and the brothel owners give the police money to look the other way. And that’s the reason they don’t really do anything. And there’s another statement which you can see at the, bottom of the slide, which is from a girl who somehow managed to run away from the brothel and then went to a police officer and she said, I’ve been taken and I’m being tortured and I’m pushed into prostitution please help me. And the police officer took her back to the same brothel and told them to put a leash on her and this was primarily because they are getting paid as such very hard for these girls to get out of this mess once they are in it.

Child Trafficking in Varanasi, India

So that’s like, the whole problem statement, and you can see that it’s a very severe problem.

There’s a organization called Guria India work in India who are using our software, who have been trying to fight the problem of child trafficking, sex trafficking, and Varanasi over the past 15 years. And you can see the picture of the founders on the left, they’re husband and wife and they’ve been doing it for a very long time. During the course, they have saved around 800 girls. They were able to file around 400 cases against these human traffickers and unfortunately the conviction rate in India for these kinds of cases is very low. So only 28 of them were ever convicted. And it takes a lot longer for these cases to get to a point, but they were able to at least restrict their bails and keep them into the police custody, which on its own safe guards a lot of the girls, because they would get out, they would kidnap more people.

So Guria’s strategy is two fold.

Guria’s Prevention Strategy: Education

One is that they go and they do their key operations, and they try to retrieve the girls from the brothels, but also the primary reason, as I mentioned before, is that the locals are not very aware or if they’re aware, they’re not, they don’t take any action towards this problem. So Guria’s strategy, one strategy is that they educate the locals towards this very severe problems. You can see on the right side, there’s a image where the volunteer from the organization is trying to educate the locals about this issue. And the second is once the girls are retrieved, they don’t want them to go back to prostitution because then they won’t have any means of earning because they were pushed into the brothels when they were very young, so their education is very limited. So they have opened these schools right in the middle of the red light district, where the girls who are retrieved can go and get education, so that they are capable enough to go and earn a living for themselves so that they don’t actually have to go back doing prostitution again.

So that’s two strategies they’re using, so that it doesn’t happen. So how do they actually go and retrieve these girls? What is the whole process? So as you can see on the top left there’s a picture that might be of a girl, which has been kidnapped and the parents take this picture and they go to the police, which you can see at the bottom. And the police wouldn’t do anything as I just explained. And then these people, these parents would go to this organization called Guria, and they will tell them, well, our girl has been kidnapped and we feel like she’s in Varanasi and then the volunteers of the non-profit wear hidden cameras on their shirts, which are usually the buttons of the shirt. And they go to these brothels and they try to sell makeup to these girls and in the process, while these girls are trying to buy makeup, their image is recorded on the cameras. And then they go back and they try to match this image to the records. If they are able to establish a match and they go to the police, they convince the police to raid. And once the raid is done they recover the girls. So you can see the images on the right where the there is image of a brothel and that information goes to the police. And then they raid and they recover the girls.

Major Challenges

They have been doing the same thing for 15, 20 years as I mentioned. The major challenges which they had with this business was that they were trying to record the image of the children with a camera which was on their shirt. A lot of the time then when would try to go and talk to these girls in the process, selling the makeup, they would capture either like a partial image of a girl or a blurry image or part of the face would be cut off or something like that and then they will have to go. Then they would come back and check the recordings and they’re like, well, this is not appropriate. And they would go back and try to do this, whole process of recording the images again. And this is a very dangerous process because the brothel owners get to know that the person who is trying to sell them the make-up seller is not actually a make-up seller but also works for the NGO who trying to retrieve the girls and all they would get in trouble. And a lot of the time their volunteers were injured while they were trying to do these operations that was one major challenge they had. The second problem was that if they’re lucky enough to capture these images properly, they would have to come back and manually match this image they have recorded to the 1500 case files they had. And that took a lot of time, and also once these girls were kidnapped the pictures of the database that happened like a while ago. And if let’s say the same girl is retrieved after a few years back, the facial features of children change drastically over time. So looking even manually at these images, it’s very hard to establish a match because the, because the girls have aged significantly in a short period of time. And also for a lot of the children, there wasn’t any image available. A lot of the parents are so poor that they never took a picture of themselves or their child ever in their whole life. So they come to these, the brothels, like are our children, our kids have been kidnapped, or a girl has been kidnapped, but we don’t really know how to describe what she looks like. And that was like another very big challenge they were facing.

ChildrEN SafEty and Recovery (CENSER) System

So this is where the AI platform, which we built came into play. So the system which we built, the first thing it did was that we took all the images of the children which they had and we digitalized it, we made it digital. As you can see on the right side, examples of few of the images from the dataset and these are all the girls which have been kidnapped. And the second problem which they had was the variation in age, in the images as compared to when they were retrieved and when they were originally kidnapped. So the system which we built had the capacity to do age Age-Invariant face recognition. So you could push an image of a child, which you can see in the image query measurement. The child has been retrieved after probably like five years or 10 years when she was initially kidnapped, but the system would be able to match that image to their younger self. So you can see that the match image in this slide looks very different from the query image, but they are the same person, and our system was able to do that. And the second case, which I mentioned was that in some of the cases, there was no image of the child.

And in that case, we built the system, which could also be kinship analysis, it could match a child to their parents, and it could do sketch analysis. So we were able to get the sketches done of these children with the help of a sketch artists by sitting with the parents. And we also have developed a system which can match the child which is the earlier age to their sketch image, which was in the dataset. And once a match is established by our system, they could take this information to the police and convince them to raid. So the advantage of the system was that they were able to do this matching very quickly and the matching was accurate. And also they were able to do this matching across age, which they were not able to do before. It’s now we walk you through the parts of the system.

Dataset Digitization

So this had the first major thing, the major challenges they had was that the dataset was non digital. So we made dataset digital, so we took almost 800 records and we scanned them. Now you can see that the images, which are like examples from the dataset, they are like visually very bad. Some of the images are, very low resolution, there’s stuff written on these images. Some of them are deviated due to water damage. And none of these images are taken by photographers who have this tendency of making the image very smooth to look more appealing, but that kills the features facial features at all. And this causes another challenge that since the facial features almost destroyed building face recognition system, which works well and work we across age was a challenging task.

Artificially Aged Dataset

And also for a lot of these cases, we only got one image. So the system which you were trying to build would have had the capacity to do age in-variant face recognition. So I really wanted more than one image for a person taken at different ages so that the system could learn how the facial features for a person changed over time, and then could predict how the person would look few years later in the age. But since that was not possible for a lot of cases, we artificially aged the data set using generative artificially aged models. And so you can see examples of that on the right where we had like the input image and we produced like artificially aged images of people and these images were used to generate agent variant face recognition. So on this slide, you can see examples of the sketch dataset. And as I mentioned that a lot of the parents were not able to take, did not take pictures of their children because of financial reasons. And so we were able to sit with the parents using a sketch artist and they were able to generate sketches of the children, which you can see on the right. And the idea was once a child is found after a few years, we should be able to push the child’s image through the system. And the system should be able to match that picture with a sketch so that we would know who the person is. Now in order to build this system robustly, we also aged the sketches using generative artificially model models because we also wanted the network to also learn how the sketch would have aged or time, and that would produce an effective system. So you can see the examples in the lower column, the second column, of course second row of the sketches, which are artificially aged using GAN model.

Kinship Dataset

And another component of the project was that in some of the cases where the image of the child was not present, we wanted to match the query image to the image of the either siblings or father or mother. So we also recorded or collected a lot of the data of the original children, which aged over time and also the corresponding sibling father and mother pictures. And that was also used by our system to come up with a kinship analysis system to which if you push the image of a child after a few years, it would match to either the father or mother, which was later used to establish the identity of the found child.

So we collected three types of data sets, we collected a data set where we have images of children at different ages.


We collected a sketch dataset, which had images of people at different ages, which were in the form of sketches and we also collected images of children along with their parents. And all that information was fed into the CENSER AI pipeline. And essentially this pipeline is constructed of an MTCNN network which is primarily a face detector. As you can see on the left that if your push a frame through it will you collect all the faces in the frame and then those faces are passed through this hybrid AI pipeline, which eventually would recognize who the person is. So if you wanna know the details of the system, you can actually look at the paper at the bottom, which is the paper which we published last year in ICCV, and it has all the details of the system. But one thing I would mention over here is that since we only had one or at max like few images of these children, it would normally be very difficult to train a deep learning pipeline because you would have to train a lot of parameters and the generalization would be very poor if you have only few images for a person.


Also it would be very hard to train since the number of examples are few as well. So in order to train the network effectively, instead of actually using an end to end supervised network, we came up with a network, which was a hybrid network where the front end of the network was fixed. And then the training was performed only in the later layers and this front end which is a fixed frontend is called Scat Net, which you can see in the image, which is a battery of Phablet filters, which you if apply to the image would give you a general enough edge to presentation, which further can be used by the trainable layers of the network alone for some specific high level features, which could be used to do the face recognition effectively. And since the number of parameters, which we have to learn are less since we are extracting already the edges, which correspond to the only layers of the deep network, the number of images which we needed to train this network effectively was less. And that’s the reason that we’re able to play in this pipeline and get good enough recognition accuracy only with a very small data set of.

Fast Convergence

And as I said, that the initial battery of filters, which are there, they extract the edges. And since you start to build high level features on top of edges, which are already extracted, you are able to learn the parameters quickly as well with fewer number of training samples. So the curves which are there shows you that the network would convert quickly, which you see in the red, if you use a scat net front end, but if you train a network from scratch, the network would converge slowly and also like the conversion would more or less be equal, but you would have to require more number of training samples to train this network. So there were two advantages of this network that you were able to train this quickly. And also you were able to use fewer number of training samples to get convergence, which was very important in this case, since we didn’t have a lot of training samples and we couldn’t collect more as well.

So now I want to let go into how the NGO uses our system.

So essentially all the volunteers of the NGO have our app, which is a sensor of which we have built for them. And there are two use cases of this app. The first one is that if there’s a child is found at some railway station or Bustan. The database take the picture of the child, they push it to our CENSER app, and it is able to match it to the dataset.


And it tells either if there’s a match or not, and if there’s a match they can go and reunite the child to that specific family. And the second is that they have these spy cameras, which you see on the bottom left of the slide. And when they go and they try to go and sell the makeup to the girls in the brothels, there is a live feed. It comes from those cameras through the system, and they analyze the faces in real time. And if there is a match, they can quickly go and take this information to the police and convince them to raid the place. And the advantage of this real time feed processing is that the brothel owners don’t have enough time to move the girls to a different location and there are a higher chance of retrieving the girls, because previously they were taking a lot longer to match these images, it was difficult. But also during that time, the brothel owners would regularly move the girls in and that was another challenge, with this system addresse.

Field Test – 1

And this is like one example of the field test, which we have done. Essentially we want to do age and face recognition. And these live feeds, which are coming from these brothels. So if you see on the top, right, you can see the image of the younger image of a person who’s walking. We just covered with a red box in the center. And if you look carefully, there’s a very big age difference. The person in the image at the top is probably like age 10 and the person who’s walking in the crowd is probably above 45 years of age. And the system is able to do this age variance face recognition in this very dense crowds, with lot of variation in terms of light and head and posture, angle of orientation and whatnot, and the system still works.

And this is another example of the system.

And you can see at the top, the image in the databases of this child, which is probably like three years of age or two years of age and we are matching it to this lady, which is on the bottom left and the system was able to match. And the system is able to identify and recognize that person. And as you can see from the scene, it’s like very complex. They’re almost around 300 upwards of 300 people in this crowd. And they’re wearing different kinds of head coverings. The distance of each person from the camera is really more than 50 feet. And it’s a very complex scene, and this is like the kind of scenes we are dealing with draw hosts, and still our system is able to perform on these kinds of scenes and do accurate face recognition.

So far the system has been used in coding operations by the nonprofit and they have been able to find one child so far.

Now you might think, why are the numbers so low? Why is the recognition rate so low? So the recognition rate is so low, primarily because the number of images with GB, which we have in the dataset is, is around 800 while the number of people which are in the brothels are way above that. So we only have a very small sample to match to. And the second is they have done only towed in search operations. And if they will do more search operations, this number might increase. And the third thing is that the kids or the children in these brothels they come from all over India, but the database, which they’re building, they’re building only for the people in a specific State. So that puts a limit, the capacity of the system and the ability to produce these accurate matches. So the non-profit is expanding on the data set side. We are trying to collect a much bigger data set, which would auditing to the hatred of this software and are also trying to do more operations, which we are hoping will improve, the hit of the software. And we hope that once that happens, we hopefully would be able to save more children from these brothels.

Thank you very much. And this is our information. So we have an office in San Francisco where we primarily do our operations and then primary production. And then we have two offices in India. If you wanna know more about us or more about the project, feel free to email me and I’ll be happy to assist you.

Contact Information

feel free to email me and I’ll be happy to assist you.

Watch more Spark + AI sessions here
Try Databricks for free
« back
Amarjot Singh
About Amarjot Singh

Skylark Labs LLC.

Dr Amarjot Singh is the founder and chief executive officer (CEO) of Skylark Labs LLC., and a research fellow at Stanford University and US Defense. Singh received his doctorate in Artificial Intelligence from the University of Cambridge, the UK in 2018. He has also been associated with the Massachusetts Institute of Technology (MIT), Harvard University, National University of Singapore (NUS), INRIA Sophia Antipolis, France, University of Bonn, Germany, Simon Fraser University, Canada. Singh has made several breakthrough contributions with over 50 international journal and conference publications. His research has been covered by BBC World News, London, Discovery Channel, Inc.