 And let me introduce the next speaker at the same time, so Kate Rose talking about combat surveillance and sartorial hacking. So big round of applause, thank you. Cool, I think we're getting started a little bit early here. Thank you so much for everybody for coming. I am Kate and I welcome you to adversarial fashion, sartorial hacking to combat surveillance. In my full-time job, I actually work to help organizations that do social good work to get the kind of technology and security resources that they need day to day. In my part-time life, I actually spend time designing and selling novelty fabrics that usually have things on them that are for kids like manatees or, you know, kittens playing around quantum teleportation diagrams and of course pillows that look like cuts of meat. So weird prints, definitely no stranger. Today's talk as a result is going to be a sample of the surveillance systems that utilize computer vision technology. Some art projects that other artists have come up with in order to counter or interfere with or engage with these systems and then you are going to get a full end-to-end tutorial on how you can do this yourself and design them on any kind of fabric or other application we call these surface designs to put on just about anything that you want. So buckle in, we're going to get going. I want you to mostly take away from this that is actually never been easier for all of these amazing designs that we read about and see online and even here in talks here at DEF CON to actually come out of the computer and into the real world and so we can put them people's hands a lot easier than we think we can. So we have our problem set which is that surveillance technology is everywhere, it is prolific, you've probably been in a couple of other talks that tell you how voluminous it is. Often that technology is implemented by folks who are themselves not AI experts, they are using stuff that's been quickly churned out, it's implemented in strange ways and in places that you wouldn't really think is a great idea. I think the big thing that I often have a huge issue with is that they generate these massive, kept forever, like highly sensitive PII-laden databases that are themselves sensitive to attack. So we can see an example here of how there was a border agency that was taking like photos of license plates and even of people at the border all breached. So we're going to go over a little example of kind of how these systems work just as a little background. I'm going to start with one that I think folks are very familiar with from the space which is facial recognitions, how does that work? You have an input image, it's usually a large field and there's a human being in it, it's broken down into quadrants and decided which quadrant or little section of the photo has something that is likely to be a person in it. It detects and puts a bounding box around it. So the idea is that it's going to crop in and then start running it through that fabulous deep neural network. This is an amazing sort of diagram from Open Face which is one of the libraries that uses this technology and then it runs it through these sort of other measurements. So we're looking at clustering, similarity detection, classification against a training set or sometimes in case it's against an identification set. So you can identify who that is right there. Thanks, Sylvester. Unfortunately there are some unintended consequences with the way that that particular type of methodology of how it goes through the technology works. There is a system in some cities in China that's implemented where if you are a J walker it uses facial recognition to try and identify and then shame you on giant billboards and then issue tickets occasionally. Unfortunately there was a very famous business woman who herself was caught J walking but it turns out that she's actually on the side of a bunch of buses. So it actually then puts up her government ID, misspelled surname, all these kinds of things that I don't think you would really want blasted across the billboard. So now we're going to automate license plate readers which is kind of like the guts of this talk. It works a lot in the same way. So instead of being in this case it is in addition to being on streets it is also primarily on cop cars. So those have a couple of major components, the plate capture cameras which are sitting on the top and then you have a processor because this is a lot of information that's coming in all the time we'll get into how much and then there is a front end. So there's like an application that the police officers in the car can actually work with as well as store any of the data and then parse it later. So this is actually a sample. You can find all these by the way at the EFF website but this is actually a slide from the Anaheim Police Department training on how to use the resulting data and then go back through it. So you can see this blue line is actually a path that a cop car took and when it was looking for a particular plate and it can tell you like the number of plates scanned whether the target which is like the hot plate or the one that is under investigation that we're looking at. You see this like little tiny note right there where it's a little pin. That's how close you can get in terms of like figuring out kind of in what range that plate was spotted. So the big problem with this kind of technology is that unfortunately it is always on and it is basically everywhere tracking absolutely everyone. They can collect thousands of plates per minute from the EFF website. They say one vendor brags that its data set includes 6.5 billion scans and grows at a rate at 120 million data points every month. In aggregate over time the data can reveal a vehicle's historical travel. Those of you who were in Senator Wyden's talk the other day on cell phone companies and how they kind of keep all of this stuff forever and that it's just so voluminous and it gives this like intense tracking of like point by point of your day often your license plate can do the same thing. So just as dangerous to keep around forever with kind of no limitations whatsoever on its use. Unfortunately there is another unintended similar consequence. I live near Los Angeles. There are lots of malls that are actually in the LA area that have this technology that when you walk into the mall you get your little ticket and then when you come back to your car after you're done shopping you put it into the machine and it tells you where you're parked. Wow it's really convenient. Except that it's doing that because there is a license plate reader trained on absolutely every single spot and they marked your license plate when you came in. You have no way of knowing this but that data was then also packaged and sold to one of the vendors for ICE. So you don't have any way to know that where you're driving in places and you're just like using a convenience technology that involves your license plate who is actually running packaging and then reselling that and using it to often terrorize your neighbors. So there are two major methods for confounding the technology. I think there's probably more but I like bucketing them into these two. You can either block the collection of information or you can overload it with additional information. So when we are trying to prevent those scanners from actually reading in the correct kind of diagram of your face or your body or something different and in another one we're trying to replicate that detection over and over and over again. We'll actually go through these two examples. So we're actually going to focus a little bit more afterwards on obscuring with additional data. This is a Calvin and Hob strip. I'm going to just read it because it kind of gets to the point. Calvin says, I'm filling out a reader survey for chewing magazine. See, they ask how much money I spent on gum each week. So I wrote $500. For my age I put 43. And when they asked me what my favorite flavor is, I wrote garlic and curry. And Hob says, this magazine should have some amusing ads soon. And he says, I love messing with data. So artists have traditionally over the last couple of decades and more recently used both methods to especially mess with digital surveillance technology. So drones when they are seeking out targets they often look for a heat register of a person. So there's a fashion designer who has actually built anti drone camouflage apparel that is meant to block that heat signature using a reflective fabric. This one is a really cool example. A technologist and artist Adam Harvey worked with a bunch of makeup artists and came up with this thing called CV dazzle or computer vision dazzle. It's named after when people would, or I think back in World War II I believe they would put on the boats this sort of like wild black and white pattern that at a distance would kind of be hard to visually interpret. So taking inspiration from that there's things that you can do to a face that actually when you draw that little bounding box around it, it makes it really hard to run it through that neural network and effectively compare and match to different faces. So this is one of the face charts that they came up with. And then I also really love sitting on the idea too of the fact that the point of this project was also to develop resources for other people to use. So they have these sort of styling tips that you can find on the website. And then they also used face charts which are diagrams that makeup artists actually use to share styles, convey concepts and teach one another. So that him and a bunch of these makeup artist friends they developed an entire library of resources for you to use. He also then developed these stickers which you can see on this gentleman which can then just be kind of placed on your face instead of having to worry about having a full makeup kit which I think is an important accessibility note. So this one is pretty cool too. There are a few scientists in Belgium that have been working with those image recognition models that not only try to tell where your face is but also is this entire object a person. This little sticker or tile that you can see on this other person is meant to sort of break up the form of the body. It makes the image confusing. So if I have something that looks like, you know, when I'm making those little quadrants or those little squares as the computer vision technology that doesn't really look like a person. It's just kind of a floating torso and some legs. But as you can imagine the tile is kind of challenging because it only really works in one direction. This one I think is super funny. So instead of worrying about your face being recognized a artist named Leo Savaggio decided that you can just use his face. Go ahead. The printing files are online. You can just go ahead and take them, stick it on and it does work. It absolutely does read as a gentleman. It's pretty cool. It's called URME surveillance is the project. So that's been touring. And then back to Adam Harvey. Man, he's working over time. He developed this textile pattern called hyperface. So a lot of times when we're looking at those facial recognition images it's looking for some of these like density or suggestion of dimension. So you have like these little spots that might be a little like eye depressions, face and nose. The shadowing can be very important, often more so than like the little stick figure smiley. And you can see here that trying to basically replicate that in different sizes as many times as we can is we're trying to distract and get those bounding boxes around each of those parts of the image. Okay. So now we're sort of like from hyperface turning the corner into additional use of that overload of information. An artist named Simone Nikhil decided to basically use shirts that replicate people whose faces are very, very often found in data sets because they're celebrities. So in Facebook it kind of overloads the auto tagging function. And now my stuff. So this is a t-shirt that was developed to basically trigger automated license plate readers. We have basically open ale PR and then a couple of other commercial plate readers were used to make this. Here it is on an actual human. You can see that it reads in as a variety of different plates and then you have it also on the back. This particular API also tries to give you some information on what it thinks the vehicle orientation is and the color, which is kind of interesting. So apparently it's an SUV crossover. So you can see also I have this dress on right now. It is actually the text of the Fourth Amendment and it works pretty well. So we're going to try to get into, thank you. I hope this is visible, but this is a live video feed on one of the commercial applications called easy ale PR. You can see those bounding boxes lighting up. Excellent. So yeah, you can see also the captured number here is kind of interesting. The number of plates that it seems, there is the dress I have on right now. Lighten up like a Christmas tree. It really loves it. So some things just really seem to kind of scratch the itch for certain APIs. And we'll talk a little bit about why. And just a quick time check. I think we got the half sign, but I believe this is a 50 minute talk. Okay, just checking. Thank you. All right, cool. I was about to say I'm very frightened. I have a lot of slides left. So you can see it works on video feed. Very excited about that. And we're going to talk a little bit about why I used commercial apps and awesome how you can design your own. Some of you might be saying that looks super cool. I love it. Others might be saying I think I could do better than that. And you know what? The rest of this talk is for both of you because now you're going to learn how you can do that because art is for everyone. So you can pick your recognition system, anyone you like. You can try ALPRs as well. You can go for facial recognition. You can go for those whole people. And then you're going to look up those libraries and APIs. At the end of the talk, you're going to have a link to a bunch of these resources where you can try them. I'm also going to very much encourage you to use commercial applications. And the reasons for that are a couple. They can be found everywhere. You can just pop open Snapchat or Apple's facial recognition in the photo function. Easy ALPR. There are so many. And it's an easy way to check your work. And it kind of is a binary output though. So you're either going to get a, you know, it'll respond to you or it won't. You can't really look under the hood and see those confidence scores or other markers. Apps usually also collect other tracking data. So I, you know, your mileage may vary. Please read the terms and conditions and make sure you're cool with all the stuff it's going to ingest while you use it. So I actually think this is really important for a couple of reasons, but mostly because of the way we're going to work together to design this experiment. So in the real world with government surveillance systems, I can't look under the hood. I usually am not going to get access to like even the training data sets that they used. I can't peel apart the neural network layer by layer and see the logic that it's using. I'm going to get as few outputs and success indicators as I will for commercial technology. So an ideal design experiment should be structured to be robust against those black box systems. And commercial applications can actually give us some of the similar effect that we're going to have. Something that I also found out even just in the course of this time at DEF CON is that if they are very similar, it is for a very good reason. And that's because often commercially sold applications that are used by state agencies have training data sets that are typically not statistically significant really from public data sets that are going to be online and used by some of these open systems. So I'm going to have you start then with your high confidence images. So in the case of open ALPR, there is a benchmark folder inside the repo. You can usually go dig through and find them. Or you can kind of collect your own. I suppose it's a little bit more painful. But playrecognizer.com also has a bunch of folders of them. Start with those high confidence images and we're going to talk about what that means. So when I read a photo into the API, it's going to spit out all that interesting information back at me, including a confidence score that it is this particular plate, what the area of the, in the photo that it is. And then it's going to give me back all that information about like I think it's an SUV crossover or you know maybe it's a Jeep Wrangler. I don't know. So now you're going to test your tolerances. You have this like awesome folder of very high confidence images. You're going to make gentle modifications to the source images to test the tolerance of those systems. I'm basically holding up a bunch of cards to the API and I'm asking it, what do you see? So shift the orientation. Change the sizing. Add and subtract parts of the image. You can cut them in halves. Rotate them. Slice and dice them. Change the colors and the contrast. Try also like moving them kind of sideways so that we've viewed at an angle. Some of these systems actually try to like clip and then rearrange them if they're at an angle. Some don't. You're going to want to find out. And I would say if you have this do it in whatever way is fastest for you. There are some libraries that will let you do this in Python. I actually personally have been using other kind of image modification products for I think like 10 years. So I'm just physically faster at doing it that way. But either way my preference is to just basically build up this giant stack of images and I've kept careful track of like how I've modified each one and like whether I'm trying all one state or different states and then I just batch test them. You can do this right from the command line. So just to touch briefly on what Pillow is. It's not the world's like most feature rich but it does have some things that I do recommend playing with and one of them is the dithering function. It just all it does is take an image and kind of guess at turning it into a black and white only image. And that is very very effective for kind of getting some of that feedback that you want in a decent amount of time. So let's take a look at kind of what I did. So here we have the license plate game. And this image is actually one that you'll start to see come up quite a bit just because it sort of ended up being this like stock photo for samples of different states plates. So I've run it through the API and honestly they all look pretty good. They work really well. I'm pretty confident at what they look like. And I think I don't know what the problem is with that last one down there. It's just not cooperating Wyoming. So now I've desaturated the image and you can see where a couple of dropped out. They're kind of not as confident as they used to be. And I'm going to now kind of run it through what I thought was that sort of dithering function. And you can see a couple more drop off as well. Also the images are kind of noisy. There's obviously like a lot that you're not going to get out of just like straight up converting an image to only black and white. So my process after that was sort of just to like keep carving out chunks of that noise. The thing that like makes it look messy and not designing. And see how often it still works. So this is an example of me testing my tolerances here. So I have this Nebraska plate. I piled up a bunch of images and I did weird different things to them. I decided to change the font. Can I use the license plate font I found online? Can I cut up the letters? Can I rearrange them? Can I change the spacing? Can I get rid of like any suggestion of an outline around the box? What's going to work and what won't? So you can kind of look at these, guess in your head which ones are going to work. That wasn't what I expected. I'll be honest. So it can be very, very challenging. Especially I'm going to have you note the top one where the spacing is different from the bottom one. The hint, both of those aren't officially the way Nebraska is supposed to actually arrange its plates. So you might get some feedback that surprises you. But as a result, you're going to come up with these, you know, after holding up your cards to your image and having all these different questions that you ask, you're going to get a really awesome set of answers. So I like to call these Q attributes. You start to get a gut sense of like what matters to this recognition system versus what does not. And often what features have disproportionate impact on whether it's a confident. So you can see the little screws here. Those like sort of design attributes that are like typical to a state plate. So they're like livery or die in New Hampshire. Even if I carve out a bunch of it, it kind of expects to see something up there. And then the distance between the letters, that's also going to vary based on which state you're looking at. And some other stuff. So there's like on some state plates you have like registration stickers and other little suggestions of images. I actually carved out a much more complicated image that was in between those letters and replaced it with this little dash. Still worked. So I advise trying that as well. I think these are actually really important because what we're doing here also is developing these rules of thumb that you don't actually have to be, you know, you have to be a coder to replicate them and use them wisely. I think we want to be able to communicate to people in other disciplines like what matters and what doesn't. I love working also with like people across lots of different spectrums of talent, both in tech and art. And I think just developing a better vocabulary for us to talk to each other is a really, really big part of this exercise. So plot enough of your images and you get a distribution of what seems to work. So here was mine. This is just what I found helpful. You might get, I think there's probably going to be a couple more axes if you're using like a facial recognition system. But you have either that your font and your design elements are hyper accurate. It looks a lot like a plate, like it lives in a little rectangle that the image expects to see. Or it doesn't. Or the design elements are not accurate. Obviously the ones that work really well are both accurate in both dimensions. And it's basically a photo of a plate. And I think that's super boring. So here we have some other things that actually still seem to work. So there was that state plate print I showed you where the font and design elements are more accurate. And it works pretty well. Even though there's like no little rectangles for things to live in. It doesn't look a lot like a plate. And then this fourth amendment print that I'm wearing right now, these all look a lot like a plate. But the font is inaccurate at all actually. And it is missing a couple of other design elements that it would expect to see. Still works. The awesome thing about figuring this out with whatever system you're using is those are awesome spaces to make art. You can really have a lot of fun. And that's where you learn that you have wiggle room to introduce design elements that you still think are kind of cool looking or that people would want to wear. So I wanted to touch on this because I get this question a lot. Which is why don't you just make your own plates that work really well? So let's look at the EU license plate standards to my friends from Europe. I can't wait to see what you do with this tutorial because you are going to have a much easier time. There are like there's amazing diagrams online of what it's supposed to look like by law. It defines very specific things like space between the characters, the width of the stroke, which font. It even has like a sort of order of operations of what the different letters and numbers mean versus the United States. So here's a list of just one state, Idaho, and it's coded by county and they're all different and it's only changes if the previous one all of them have been used. So that's one state. I made a little meme for this, which is from Ron Swanson from Parks and Rec where he says not to worry, I have a permit. And he just hands a piece of paper and just says I can do what I want. So, America. All right. So these are some of the things I ran into that did slow me down. They were challenging so just expect them and I think you'll have a good time. You have to balance whether or not it looks like a plate especially. I mean getting rid of that rectangle, it just doesn't like it. So you're going to have to struggle against that a little bit. Other desired design attributes, obviously I don't want to like have it always looks like just a pile of plates. It's worth the effort but I think you're just going to spend a little more time on it. Some states fonts and spacing don't work in other state plate formats and you might find yourself suddenly trying to like mix and match a little too much and then suddenly everything stops working on one side of the image and you're like dang it. Sometimes the API will also, as we saw in my example of the ones that are all modified in different ways, allow another state plate's lettering format to fake out what it expects to see for that state. As a result I've gotten ones where like it's clearly a plate that says Missouri on it but the API loves to tell me it's from Florida. I can't make a change. I can't figure out why. So deciding, I'm going to also spend a little time on this one which is deciding where to ethically source your plates, faces or any other images that you're going to use. So I did my best to find what I guess I would call out of use images. So things that are like functionally public, they're kind of like a stock image or something that's been reused a lot or it's on Wikipedia or you know it's just functionally something that it's so public that like you adding that to the set isn't really going to change much or they're functionally out of use. So you can go to junk yards and buy old ones. You could find them hanging on a wall of a restaurant or on eBay or on Craft Decor. So they're things that are people are not driving around. I also get the question a lot which is why don't you try to figure out whether or not they do belong to somebody? Because that costs a lot of money. Unfortunately, it is one of those things where like you're not supposed to be able to pull that information just from any plate on the street. If you do have enough money, anything is available to you, I do not have enough money. So I just have to do what is within my power as an artist and a good person and somebody who's trying very hard to like be a good example to do what I can without either being a millionaire or otherwise having access to perfect information. So please try your best. Also that is to say if you want to put your all in play all over a t-shirt, that's your decision. You know, might be then spotted around some other parts of town and you might find that beneficial. But I'm not the boss of you. So here's another amazing sample solution. Look at all these wonderful people's faces. Wouldn't they be fun to use on some kind of facial recognition project? Guess what? You can because none of these people exist. They are from uh, this person does not exist dot com. Awesome. Sort of generative adversarial network. I'm gonna flip back for a second. That just makes people and they don't look like or they're nobody. They might look like somebody. So just, you know, model that out as best you can. So now I have my images that work. I figured out I have this great pile. I want to put them on fabric. They look awesome. Anybody can design a pattern. I really truly believe everybody in this room would be capable. Like you may not think of yourself as like an art person or a fashion person but I'm so glad that you are in here because you are absolutely have all the skills at your disposal. Um, so we have a front to back. This is like the rest of this talk is gonna be like how do I deal with all of the stuff that I have and then making it to the point where it's useful because I've read a ton of papers on amazing adversarial patterns and pictures and all these things that work. It's really hard to find out how to access them. And I think everybody in this room could also be part of that change on how we put those in people's hands. So basics, uh, the half brick is the thing that you're going to see most often. If you have a bunch of like things that work, just putting them together in a line, uh, then moving it down one and shifting it over about halfway. Uh, gives you a pretty aesthetic pattern. Works pretty well. Uh, the diamond repeat is actually the basis of most seamless pattern making. You might actually want to like make note of this not just for fabric but I think any surface design. If you're thinking about it in your head, I'd love to put this on a car wrap or a mural or something interesting. You're going to have to know this. Uh, so I've changed the contrast a little bit so you can kind of better see what I'm doing here. Um, so you have these tiles that repeat. Uh, it's because each individual one, like I made this cool design. Let's imagine it's like a cluster of plates or something. I'm going to divide it into quadrants and then move the one from that side down to like, so if it's the top left, you're going to move it down to the bottom right and complete that with every single quadrant. And that is how you get, it's really hard to tell actually, but I have a plate cluster and a circuit cluster. And all I did was I chopped it into quadrants and then I moved them down each corner and then I filled in the middles and then I got a beautiful seamless print that looks pretty nifty. So you can say, uh, it sounds really hard and like it involves math. It does. I'm sorry. So here I'm going to try to give you some examples of things that don't need math. Uh, so here is again our, uh, our glamoflage, uh, facial type of t-shirt. Um, what they had done, the artist is to just create these like horizontal blocks and then they're a little bit different, the three of them, and they stacked them, staggered one on top of the other. Uh, this is really, really helpful because if you just go big enough, you could just slap a t-shirt right on it. It's really okay. You don't have to worry about whether it repeats and whether the edges fit together. So, uh, the contrast in the images that you're using for your facial recognition patterns are very, very important and then if you're like, oh, but I really like these colors and I want to make them work for my design. Uh, I think one of the things that's going to be really important to you then is to learn just a tiny little bit about color theory. I promise that this won't take very long. Uh, but tint shades and tones, so things that are mixed with white, mixed with black, and mixed with gray. Those are going to affect your contrast of your final image, so pay attention. Uh, things that look good together tend to be all the same tint tone shade levels. Uh, you're also going to find that when you're running this through image, uh, sort of image like editing software, um, you're going to want to like convert that to black and white once in a while and make sure you check, uh, the actual technology is sometimes sensitive to that. So that's why the top looks great and the bottom, not so great. Uh, you can actually then, if you're like, this sounds super hard, you can extract different colors like you have this DEF CON art, uh, from last year. Uh, those are part of color families, the things that are all next to each other on the color wheel. Uh, we have a really cool EFF t-shirt from last year which uses triads, which are two things next to each other and one opposite also tends to look very cool. Uh, and borrowing is okay. I think like getting over the hump of like, I'm not good enough at this to make something that people would want to wear, just grab an image that looks aesthetic or, uh, a piece of clothing in your closet you already like and you can run it through some of these. I have the link list at the end of, uh, you know, like Adobe color or cooler or degrave.com you can use to then extract samples. Monochromatic is also super cool. We're all hackers and we all like wearing black. Easy enough. So let's put it all together. My example here, uh, we have our fake faces. I'm gonna do some complexity reduction. Pretty simple. Not gonna spend a ton of time on it and then do our derived color combination from some, uh, other prior art. So I now have this facial recognition triggering jacket. All I did was just replace some of the colors and kind of strip out a little bit of the noise. And yes, it works. It doesn't have to be so hard. Uh, so here we have the Apple's facial recognition on the image, on the jacket. It does actually do what it's supposed to. You could continue reducing noise and improving simplicity until it stops working. And then you're gonna get that information back on how you can make a pretty snappy pattern. So from pattern to production, this part's super important because I think we all have lots of great ideas, but getting it actually out there is a really big hurdle and perhaps been previously. Uh, fabric is not pixel perfect. I really love the whole like turtle, you know, AI thinks turtle is a gun project where it 3D printed a turtle that the pixels had moved around such that it, uh, interfered with the model to the point where it thought it was a weapon. Uh, but full adversarial models tend to need to be pixel perfect and they only work in one direction. Uh, soft goods have a lot of constraints that way. Set your expectations to wearable clothing with reasonably durable imagery. So what is reasonably durable? You have to think about a wraparound pattern like this one goes all the way around. Uh, does it work with the curvature of a body? So I am a plus sized person. I had to size up to make sure that things were not warped when they were either too small or fitting in a strange way. Um, as a result can you pick fabric or materials that are going to help you out in that endeavor? Should your garments be loose or structured? Uh, this is a picture of a raincoat. Might that actually be stiffer and be more reliable when I'm kind of at different angles or sitting down or standing up? Uh, you're going to want to test and test over and over again. I'm going to get into some of the ways that you can actually develop prototypes but I think each individual step is important. Flat digital prototype. Test it. If you're using a system that lets you do a digital mock-up like this one, go ahead and run it through. I know it's not a photo on a real person but it's helpful. Uh, and then you get your print prototype made. Test it again. They are often really good creator discounts on a lot of these sites and so you can order some samples for for quite cheap. So, uh, big thing here is that the digital printing process on fabric is made using a heat bonding dye injection process. It tries its best to get in between the fibers and it's not always very successful. So, uh, you're going to want to make sure you work with the texture of your fabric, not against it. Work in very high files. I've seen so many things go right because people like literally need more JPEG. Uh, so, uh, you can actually also order fabric samples. They're like $2 and they send you this giant box and it has like hex codes actually printed on them. So, uh, they they really try to help you out here. Uh, so it's easier and cheaper than ever to make them yourself. They only make each one you don't have to hold inventory, you don't have to buy 25 shirts and keep them in your closet and hope somebody buys every single one of them. They are made to order. Uh, it is very low risk. Uh, you can print on demand. You can print samples very rapidly. I have vendors that I've worked with that I've really liked, uh, printful. They have a really awesome way that you can get this dress right here, a little bit more expensive, but very nice quality. Uh, spoon flower where they can make fabric by the yard in like three different ways. They can do wallpaper, which is pretty cool. They can do gift wrap in case you like giving weird, nerdy gifts to people. Um, and then Will DeMache, which actually looms sweaters, have a little process way down. Before it used to be that you would have to like be an academic at like an institute, and you have to get a budget to like make sure that you could actually afford to, you know, print these things out. Um, 15, 20 bucks. It really isn't that bad. So I think it's also really important to bring the final garment cost down because people will take a chance on that concept clothing if they want to. Uh, printful shop that's a provider for all over printing in Los Angeles. Uh, so it actually also might not take as long to get here as you think it will. So manufacturing your own. As you wander into the world of drop ship, please, I had, I just felt obligated to put the slide in here. Please don't get scammed. Uh, please read the reviews. There are even on a creative idea that you think the world needs to see. I think that's really, really important and it's never been easier to do that. So you're going to balancing your manufacture quality, distance, shipping time to your happy customers and the wholesale price. This is really important. I was so inspired so there's a, uh, indigenous artist named Mona Cliff who's been working on putting QR codes into traditional beading. Um, it's still made methods. It does not have to be digital. Still works in these recognition systems, still can be read by, uh, machines. So the objective here just to kind of like sum that all up is we're lowering the barrier to entry on caring about surveillance. Everybody wears clothes ostensibly. Understanding how the technology works is hard. Buying clothes and wearing them is easy. Something you can engage with in a tactile way makes it fun how they should care about this thing. This is actually an anti-Paparazzi scarf when you try to hit it with a flash it blows out the photo. Very cool. And if we want people to care about surveillance I really feel like we're going to have to keep doing this in lots of areas. Um, you literally have to give people something they can hold in their hands and help them like understand and use it and they could try it themselves and really get a sense of how to test themselves. Be little citizen scientists and then start seeing that everywhere and I think clothes can be a great way to do that. So just to give a little sprinkling of of course because this is the crypto village of other things that you could possibly think about if like triggering recognition systems is not your game. Uh, there are historical analog applications for fabric. This is actually a really cool thing but spies would have them either in the lines of their jacket or they would tie their hair up with them or put them in their knitting bags so that you could pull them out and they were quiet. They're printed on silk they're reasonably very durable and then they would say like okay I've used this letter transfer block this particular cipher and they would put pinpricks in them or embroidery in them to denote that they had like you know letter substitution like a substitution cipher put into fabrics and woven in they can be done big or small pretty cool and I wanted to put in those slides because I'd like to get thinking at the end of this talk sort of about how do we weave anti-surveillance design into the world. I think one of the big things that would have helped me a lot was how to work yourself maybe you watch this whole talk and decide I'm not very good at art and this all sounds very confusing and I'm not a fashionista but do you know somebody else who is who could maybe unleash a lot of creativity could you make them a tool that lets them not have to use code to take these systems and engage with them I think that's honestly the biggest thing that you know Android or on mobile devices in particular is a really awesome fast way to check your work so I think also we want to think about bigger not just people I think surfaces design can be anywhere I would absolutely love to have a mural up somewhere either in like Metro Los Angeles or any major city where cop cars go by all the time those databases as they pass by anything that you can do to make the dataset less useful they have to pay to store it they have to pay to analyze it just I think every little bit of being non-compliant is super helpful so think about like murals, billboards, clothing lines, posters for events like a hacker conference so I really don't want to getting better even in like the cycles of automation are so short now that I can barely keep track some of the anti-surveillance patterns created by other artists as I was doing research for this talk already no longer work happens but that's the point right so hacking surveillance systems shouldn't maybe I'm going to pause it not be about trying to outrun the arms race how do we normalize mimicry aesthetics and methodology that you can repeat every time these things catch up for continuous testing and iteration it's not a one and done process I think if you remember from the movie the fifth element I really love this guy so you know Bruce Willis looks out in the hallway and he sees what he thinks is an empty hallway and it turns out it's just the guy wearing a picture on his head it works I don't know so don't limit yourself so I have all the slides up at fashion.com you can see resource links those are the slides will actually be up right after this talk you'll also have resource links everything that I used you can use please treat this as a tutorial show me what you make and then I have the available designs so you can see how they were composited on to different types of dresses and I just love being able to do that I'm a researcher Dave Moss honestly the concept couldn't have happened without him he just mentioned in passing that the specificity on these systems are low they read billboards and fences and all kinds of junk so that unlocked a lot of creativity so maybe information you share could do the same for somebody else probably privacy and that is the end of my talk thank you so much everyone I really appreciate it so we do have some time for some questions if you have a question please come line up by me oh my twitter is Kate Rose be like the insect alright no questions is okay as well about the fact that a lot of these algorithms for surveillance are also used for safety critical systems and automation they also use what sorry they're also used for safety critical systems and automation like self-driving vehicles oh that's really really interesting I think one of the things just to repeat a little bit of the question that some of these are like safety driven systems as well so I don't necessarily always want to be you found a mistake I think there's actually been some really cool research in particular of showing how like salt lines on the road there's another talk that was given on how to like red team how to red team AI without being a chump I like that title and they had talked about the fact that if you salt the road in particular ways that messes up lane detection right to the company tell them what you found treat it like any other kind of exploit I actually think one of the things that kind of stinks about using license plates for so many different things is that like yeah it's it's pretty much something that anybody could do with any plate they want and I don't think as a result they should be used for highly sensitive things like incriminating you for awesome talk really like thanks did you find anything about testing particular design across multiple models and finding that you know something worked really really well in one model but like really failed with another model that you had and like what did you learn from that yeah great question so across the different open versus commercial systems that I was working on some of them are meant to be used as like a person who is maybe a lay person using their mobile device they are more sensitive they will basically take in a lot of information of anything they overvalue that like rectangular function a little bit more than anything else versus something that's kind of like reading things in as a flat image as the primary function of the API versus a video feed with a different conclusion but yes they definitely vary I think often folks and they should they frequently rewrite and make their own sort of like recognition algorithm that overvalues one thing over another maybe based on customer feedback I would love to get my hands on one that's used by state agencies apparently you can get them on ebay so if anybody wants to pitch in with me if your dress became really popular the software providers could just start filtering out all the text of the fourth amendment and then they'd never get false positives anymore so is the tooling in place that you could make it so that every dress has different text on it so it's very cool that you bring this up but it's possible yes it's one of the things that actually becomes really cool and interesting for me myself remaking those like without digital printing like remaking screens for a silk screen or remaking loom like your loom orders with a wholesaler that would be really hard now actually it's very easy so if I find that it's screened out or they just don't like my clothing anymore I can go on there dig in my pile of 20 minutes so I can outrun them for a while I think with your collective help as well if we're all coming at them from different angles it makes it a little harder so thank you great question one of the things I was thinking about immediately when you were talking about all the modifications is possibly automating that so you could just throw something at a machine at a model and then so you could quickly find something that was aesthetically pleasing that might look cool do you think that's reasonable or possible or how much of it was like you as a human had to figure out this would be a reasonable transformation I think it's super automatable honestly so some of even the technology that you were looking at that extracts pallets from different things kind of operates on functions that I imagine would be like most prolific web developer or else like a tool would be amazing but if people wanted to collaborate on that I'd be super open to it all the methods of things I was doing are if I use them in adobe it's just because my hands are faster than like literally figuring out how to code that image modification tool but it's very possible I would imagine that slider line that you just drop a bunch of icons in drag them around and it makes the repeat tile for you so it is some of it to some extent actually is already out there but yeah I think it would be pretty easy over time to work on that I know that I'm increasingly also really interested in like generative adversarial network says one tool to address that so instead of being processed by a robot not being in the manual processing and your face will be very carefully recognized wouldn't that happen I'm sorry I didn't quite catch the question so that that's a threat right so is there well how probable would it be that you are since you are broadcasting well distance then you would be screened manually making repeat yourself yes so what I'm saying wearing these manifests that you are well resistance therefore the automated system can flag you for manual processing so yeah that's true I think like honestly that that is just like one of the risk points that you have to incur when you're kind of working on this sort of stuff is that you know you can have I think you know the fact that plates are ingested so rapidly and these databases are gigantic and messy and implemented by people who are generally speaking not that they're not that thoughtful I really like don't mean to not flatter anybody but at the end of the day it's a tool to get a different job done and I think like the actual