Closing sales online for Animal science

How to create outlook signature

thank you so this is joint works mandame will give the meat of the presentation for sure let me see it this doesn't show up on here okay so um as we're talking about the trade is vast and it seems to get bastard by the day in part because it's just the more we seem to learn the worse things seem to get so this was the first study published on the sea snake trade for instance have the Gulf of Thailand into the Chinese market a couple years ago and in some ways I think that there are times we have to work on demand and supply of course but there may be times where we may choose to focus more on demand or more on supply depending on how diffuse or concentrated those two things are so we're working really largely on the demand side or closer to the end user by looking at people buying wildlife online and there's attention because we know President Obama has made this great executive order and would like to reduce the trade in illegal wildlife and at the same time we have people wanting to connect the world more and more and this being the mission of course private companies but also governments and the idea that connection is going to make the world a better place that may not be true for everyone especially if you're a Pangolin for instance so this is just you know sort of saturation of Internet users by country even though we have more internet users in China for instance there will be increasingly more and more and what happens when these markets go online what happens this is actually a fisherman who was out catching one of those sea snakes he's using a piece of rhino horn there as a piece of traditional medicine to combat the poison in the sea snake so what happens when he has access to online markets so what we know about online markets to date has really been provided by civil society which i think is interesting that academia has not been out in front on this issue so the International Fund for Animal Welfare has done some of the key work showing through manual searches what's available in terms of the illegal online trade they are focusing as we are on site ease appendix one species there are about six hundred and ninety nine of them these are species that have been agreed to at the study's convention to prohibit international trade in these species and Sadie's appendix 1 has been called a political listing rather than a scientific listing but in any case it's it's illegal so it's a very good metric so they looked over a six-week period at eleven countries in this first report published in 08 and then they adjusted another one in 2014 where they increased to sixteen countries over a six-week period and found yeah about 10,000 ads and 7,500 those or so we're appendix one species you can see the breakdown by species over here and they're using teams of volunteers across these countries coordinated by their central I fall office Tonya McRae steel is the head of this project oh sorry um so they look at they've looked at 280 online marketplaces but web sites like eBay like Alibaba and country specific sites like that now they're just moving into social media to as a place to observe but this it's funny to me that it's being done you know again not just by civil society but manually as well when you have the kind of computing power that you do in this room so there have been some other attempts also to look at sort of studies one trade online this was a paper I think Lucas is familiar with that didn't exactly use any computational methods but looked at 24 sellers on ebay they looked over a period of six months twice weekly at sadies one listed cacti and then they compared what they found at the point where they reached a thousand adds to the site ease permits that had been issued for those cacti so you can trade in appendix one species under exceptional circumstances and same with appendix 2 and they were able to show that a lot of this trade was indeed illegal because it didn't match up with the customs forms that had been achieved through site ease and then there's a new paper out just this year on the dark web which people are talking a lot about and its role in the wildlife trade but this study looks at about ten thousand different items found on the dark web and there was only a single instance of an illegal wildlife product which was a hallucinogenic cactus so this first sort of glimpse at least in the in the public record of the dark web doesn't look promising for the illegal wildlife trade and that may just be because the trade exists on the open web and it's quite obvious from the I fall reports for instance that that's true so we're interested in what kind of comput a computational methods have been applied to the illegal wildlife trade and the paper that that most comes to mind I think the only one we know about is this one published in 2015 looking at the shredder lines on ivory and using machine learning to then detect through a photo algorithm right ivory sales on ebay over an eight week period now on that eBay has cracked down so significantly on ivory sales these people apparently are a little bummed out that their model has been proven you know sort of obsolete so quickly no I'm sure they're thrilled I'm sure they're happy they'll use it for other things and there still is ivory online of course so then there's also a couple of computational projects that really um they kind of appear in the same vein as ours but are really quite different this is G Delta some of you may have read the paper by the the author behind this in foreign policy and also this they really apply similar methods as health map which is using a kind of machine tool to aggregate news and website information from seizures of wildlife to map the trade but this in our minds also highlights enforcement hotspots more so than the trade per se or may highlight enforcement hotspot hot spots rather than the trade so this is really looking at what happens after governments and customs have already gotten involved so you can give us a glimpse but I can't tell us about like what is going on online right now so we have two arms of our project the first is academic in nature because there hasn't been a sort of large-scale assessment of endangered wildlife online so our first contribution is a paper that's just going to assess the critically endangered endangered and vulnerable species and I you seein red list tens of thousands of species plus societies one species where they're being sold on english-speaking websites and then the second project is a subset sort of of project one which is to identify Sadie's one species that definitely or probably or maybe even with a low probability illegal wildlife sales online Sosa nandan is going to talk about how you actually do that Thanks corpus thanks Jennifer for giving such a nice introduction to the project and we're submitting our final application [Applause] so let me star dive directly into the goals of the project so what we are trying to build here the first deliverable is our customized search interface which will create a vertical slice of the web presenting results which are relevant to the wildlife related trade the second is whenever we see any animal any species in an ad identify whether that species belongs to appendix one or not in other was legal or legal or illegal so like you see this ad is it like brown bear which is listed in situs appendix 1 and the third question is identifying illegal product not related to wildlife like this one being sold as bovine bone is this ivory or not so these are the three different goals we have for this project so the proposed solution will have like an automatic crawler which will crawl will collect data automatically from the web and anything detected will be reported in real time and in in this process it will also identify potentially illegal items illegal ads or as dealing with illegal items and report those and finally generate a report which like popular trends and hot hot spots and what are the items and species more frequent than the others in this space so in order to do this we have targeted 70 different sites english-language sites which includes online marketplaces including social media like individual retailers sites where like marketplace is a platform in the internet where buyers and sellers like interact with each other in addition to that we are also looking into sites where sailors are said in the directly interacting with customers are selling stuff like taxidermy on length taxidermist order online auction houses etc and we got this list of sites based on a general web search using keywords like wildlife for sale it's a very generic keywords and looked at the domain names of the most frequent sites which were written by these queries and we also looked at sites which are selling things not like Wikipedia or dictionary.com we're also written we ignore those kind of sites so the key words we use they're basically two types of keywords on your right this is the list of species animals p animals listed in the site excited appendix 1 and in addition to the scientific name we also are using common English Spanish and French names and on the left hand side we are using some additional keywords which are basically general words describing animals like snakes birds or like big cats we are using keywords related to body parts which are being very frequently created lexical for feathers and like popular products like trophy or like taxidermy and the example I showed before like code was which we know about like ox bone or bovine bone which is used for ivory trade sorry so this data is automatically collected we have around like combining the species and the additional keywords around more than fifteen hundred thousand fifteen thousand keywords and we take one week to complete the cycle collect all the data relate all the keywords and the sites in one round so what are the features we can potentially use one is the text in ads you don't have much text but whatever text we have we can that can be useful we look at the price that can be a very good indicator of something being legal or illegal we look at this is probably not clearly visible we look at the item location and where it can be shipped to that practice that will give us an indication whether the item is crossing an international border or not and also the delivery information how fast it can be shipped I come back to layer can come back to this later why this is an important feature finally the images like images can have like has can potentially have a lot of info if information can be extracted properly so next data is clean so whenever you search for wildlife related things in the internet you get like t-shirts or software's which are not interesting to us you get like the reason image can be very useful like the text here is like the third image the image says it all that it's not a real like head bound also we look into like the price because if it's too sometimes too cheap to be a like an illegal product and finally we keep or the postings which has the potential of being illegal so what are the challenges it's not a very easy problem first of all the use the usage of key worm-like code words like move and one in this case or sometimes like sailors use the words like replica or reproduction or four which are like maybe it's not it's a real thing being advertised as so the so sometimes the text can be misleading we need to look into that and another problem we face is like the taking the case of zebras there are three different species of zebra and two of them are listed in sight is one the third one is not and whenever we see any zebra related product we see that it's the virtual zebra which is not in the site is list always that species name is mentioned so we need to identify whether it's really virtual zebra or plain zebra or not or it belongs to some the other endangered species so so what are the like the hints we can like what are the cues we can get from this adds that can be helped us to identify something is illegal or not first of all inconsistence inconsistent information they look at these two different ads both are claiming that they have a site is permit and if you look carefully it's the same permit so sometimes we need to look into this kind of information which is potentially inconsistent that might raise suspicion second is incorrect information like this ad it says its priests itís it's it's a vintage product but is it so it might be are like a very recent product advertised as a vintage piece it's-- product and finally looking into the shipping information whether it's crossing international border and finally the shipping time so this ad where I took this for screenshot from was done on May March 31st and it says that it can potentially ship this item within 23 days which is too soon and it's crossing international borders says it can be shipped from united states too many other countries so it's too soon to get the situs permit so which makes it like let you might think that it is potentially an illegal product so how can we collect all this and aggregate all this information to bidder build a model to identify illegal ads so the problem here basically becomes like building this mapping function which will map an online ad or online posting into a variable via binary variable Y where zero will indicate it's illegal and one it will one means it's a sorry zero means legal and one means illegal so this ad can be converted into a variety of features and some model parameters which will be taken as input and the output will be computed so as I said before the kind of email features we used our price text shipping information item location and the images and we can also like asks ask experts to have give us additional information like people who work in ivory trade knows just by looking at a photo of an ivory product with its ivory or not and we can just take those cues to and incorporate those into our model to identify specific products and so the first step which we tried combining these features is your animal detection to see whether like we can detect outliers from this data so look at this example this is a prize distribution of ads with the word bovine bone we see that most of the items listed as bovine boon are priced less than five thousand dollars and this is in u.s. dollars and the sort of this is the prices in u.s. dollars and the the y-axis represents the percentage so most of the items are like less than four thousand five thousand dollars but we see the certain items which are like far away from the the non-standard like the mean value of the prices similarly all the ads which had the words replica in it so 90 more than ninety-five percent of mass is below or close two thousand dollars but there is like few items which are like exceptionally priced so are these like an indication of of something being illegal which is sold as a replica but it's not so we try to generalize this this idea by combining the species names and the the body parts or the product type like say for example combining tiger with Tiger skull tiger skin Tiger taxidermy and having this like a competing a Cartesian products of species and the item types and making like clusters of this and plotting the the price for each cluster and we see for majority of the clusters the standard deviation is low but for some cases we see like outliers like which like some items which are far away from the standard range so this can be this things can be taken as like illegal products so we used a very simple outlier detection algorithm which is about like it's a graph based method where for each item you take like the clothes k closest neighbor k-nearest neighbor and draw like directed edge and if you find that some item has less than a threshold value of incoming edges that means it's far away from far away from all the all other items that means such a potentially a potential outlier so based on this we detected and de the threshold value is that user defined which we used to try to learn from the data itself we can identify outliers we used another method which is called anchor variable which is like which has been used in idea of classifying like clinical predicting clinical stage from medical records so this the nice thing about this method is it for when you have data with very noisy labels no labels or only positively leveled our data and how this works is like you try to find so if your data has two types of variables one is observable and others are latent you can't observe the values so only by looking at the observed of the observed variables you can try to find out like find out but one particular of the variables which has a very strong link to the latent variable so in that case those this particular variables call the anchor variable you can replace the you can use this observed anchor variable as a noisy label and try to predict that so you don't have to worry about the worry about the latent variable at all so how this works is like so the idea is you have to find cases where the observer area the anchor variable and the observed variable has likes the same value if the anchor variable is one the observer the latent variable is also one with a probability of 1 like sorry I put that in the in the slide so basically you have to find that like if the probability of y equal y equal to 1 given a equal to 1 should be one in that case even replace the of the latent variable with the observed variable and redefine the prediction problem this is like I victim this using like binary variables but it can be easily extended to like continuous variables so how we can use this idea into our method is like we can use price or certain code words as the anchor variables which we can see in the ad we don't know where this legal illegal and try to implement this way of classifying the ads based on what we see and so and we can use apart from price or the code words as anchor variables can also ask for very specific products of species experts what can be a good anchor variable which is very has a very high information content to predict something and so this is what we have done so far so the next steps are so we have looked at in two like very specific features of the ads or the online postings but you know now at this point we would like to look into images so we would like to like get information from the images which can give us valuable information also we are like we are into like identifying the we are also trying to understand the scale of illegal wildlife trade in the internet and for that we need to do some kind of duplication duplicate detection that you see this this same product has been advertised in two different sites so image processing like image matching can very easily help us to identify this kind of duplicates and so far we have concentrated only in English language so we would like to extend this work into other languages like Russian Chinese a certain south southeast asian languages and we are our methods are giving us some estimation of whether an ad or an online posting is legal or illegal but we need to validate those results for that we need to engage like experts or people who are in this work in this space to help us with that so before concluding I would like to thank the Muslim data science environment at NYU to for sponsoring this supporting this project and all our partners from ms are like Lucas is here we will help us r a lot I for university of washington and US Fish and Wildlife Services mr. with that I will end my talk [Applause] questions so is the goal here to build a fully automated system or is it something very raise alert for human experts to look at it later a mix of both so the tools purpose is to give some kind of probability value to an ad being legal or illegal or not and then the human expert can use that value and use his or her own judgement to take a final decision and yeah we think that some retailers they want to use it in advance of even posting me as like to supplement their own efforts some of which are next to nothing so that they can screen out ads before they even post them without a problem for it trying to crack down on this problem with no sir people see that you are identifying their and illegal they all changed it start changing their keywords and so how would you continuously I did't I'm ongoing water all the keywords that correspond to giggle items are ready when you are back to back down on the normal wave right now they may also move to the dark wait for exam if the dark wave thing happens I we would say that's a success of our project first of all and second is like that of the previous question you had like.we yet we don't have a method to identify code words we are using what the code words we already know people who have research in this area already know so we also is part of the problem or part of the project identify key code words if that's that one is done properly this moving target thing wouldn't be a problem we we should be able to identify new code words and if it moves to dark web that's bad and it like require completely different approach to identify those items it speaks to where we prioritize wildlife that it is on the open web though because of course human trafficking is not and it's so rampant in the open web like at this point this is the issue we need to tackle [Music] question he said you're collecting the data reporting reporting into coal to provide the information so yeah so the first um we had three main engines in mind which were the NGOs like I thought we thought might be sort of clients of the data and then retailers like eBay and then enforcement like fish and wildlife and those are still are three and we have they move around in our minds as to who is the highest priority or who we should tailor the the tool for so I think those are the three were thinking of at the moment and we are in conversation with all of them so it's a question in the end of who we think will will use it most and where would make the biggest difference for a wildlife you mention click does ricky the use of connectors yeah that's also a possibility individual collectors but if I won't find something like I'm a good person also I i only have information whether what I buying is legal or not that's also possibly these keys but the move I may not share anything of course much yeah I guess that makes up the question I think this is made of when the bad guys might you might use it to figure out what you are thinking about the clear so you may conveniently figure out those things that quantities is really like making it up sort of publicly available tool opens the door for vigilante justice type reactions which I mean I know we're all for a wildlife here but at some level but this should be a law enforcement issue and not vigilante groups going out and punishing like this a little lion hunter or whatever so they these companies have trusted flagger programs already so ebay craigslist no because Christmas doesn't care but you know oh you know every company does but most of these companies have trusted flag or programs and this is that's a fantastic intersection for the companies to maybe not do it themselves but to make this tool available to the public but which is kind of a sub select group of the public and and to be clear I mean you know we can say things you should be law enforcement issues that or not but you know they're generally being taken care of in the hands of private corporations which choose to or to not or not to post particular ad suit there's only very specific domains where private corporations and public law enforcement agencies have really partnered together on strong data sharing and that's principally around child ography and human trafficking all right we're going to move on to our final Hawk so thank you