Circumvent Facial Recognition With Yarn

Knitwear can protect you from a winter chill, but what if it could keep you safe from the prying eyes of Big Brother as well? [Ottilia Westerlund] decided to put her knitting skills to the test for this anti-surveillance sweater.

[Westerlund] explains that “yarn is a programable material” containing FOR loops and other similar programming concepts transmitted as knitting patterns. In the video (after the break) she also explores the history of knitting in espionage using steganography embedded in socks and other knitwear to pass intelligence in unobtrusive ways. This lead to the restriction of shipping handmade knit goods in WWII by the UK government.

Back in the modern day, [Westerlund] took the Hyperface pattern developed by the Adam Harvey and turned it into a knitting pattern. Designed to circumvent detection by Viola-Jones based facial detection systems, the pattern presents a computer vision system with a number of “faces” to distract it from covered human faces in an image. While the knitted jumper (sweater for us Americans) can confuse certain face detection systems, [Westerlund] crushes our hope of a fuzzy revolution by saying that it is unsuccessful against the increasingly prevalent neural network-based facial detection systems creeping on our day-to-day activities.

The knitting pattern is available if you want to try your hands at it, but [Westerlund] warns it’s a bit of a pain to actually implement. If you want to try knitting and tech mashup, check out this knitting clock or this software to turn 3D models into knitting patterns.

38 thoughts on “Circumvent Facial Recognition With Yarn

    1. This should be a solvable problem with utility outside of just fashion choice. Patterns like this are probably MORE useful if they appear uniformly black or are otherwise not human readable. That decreases the chance that video of you gets flagged for ‘this person is attempting to evade surveillance, so a human needs to see this and manually select their face for recognition’.

      You can take advantage of the way surveillance cameras work: because they are also used for low light, they don’t have infrared blocking filters and can see well into the infrared spectrum. That means you can make patterns you can’t see, but they can.

      Knit with two different fiber types that have low reflectivity in visible light, i.e. look black, but have different reflectivity in infrared. Look at material properties. Two examples come to mind. Acrylic is easy to cut with infrared laser cutters because it absorbs infrared well. Black acrylic should still look black in infrared. On the other hand, black polyethylene sheet is used to cover soil that needs to be warmed by the sun, as it transmits almost all infrared light. Black acrylic and black poly yarn should have minimal contrast to the human eye, but appear very different to infrared.

      I think at this point you’d want to use a knitting machine, because if the point is that YOU can’t tell the difference, the pattern is going to be hard to do by hand.

          1. I’m not sure it is any more of a horrible idea than putting a face on your shirt anyway…

            Would be an interesting experiment though, not sure quite what it would tell you – would law enforcement actually turn up at any point? If they don’t does that mean they are not using FaceID type stuff everywhere yet or just have good filters to quash false positives, and if they do show up how long does it take?

            I suspect they would never show up – it is on the shirt, so while the AI might flag it the human would quash that immediately, as surely you don’t send the squad out on every AI flagged ‘face’ without checking.

            The real adversarial move and challenge would be wearing a 1:1 scale of one of those faces printed on your ‘covid’ mask style masks. That however would be rather more likely to have tragic results, as you might well fool the human with low quality footage, so I’d not really suggest doing it.

          2. There’s some other ways to get into trouble

            1) Use your phone near a crime scene, and get caught by a Google Geofence warrant. Quick, hire a lawyer and appear in court within seven days. Possibly get thrown in jail for a week without being told why.

            2) Wear a T-shirt with cartoon bikini clad women printed on and appear in public media. Whatever career or life you had is now over.

          3. >I suspect they would never show up

            I suspect they would, to stop you messing with their system and causing false positives. Might even throw you in jail on some charge.

          4. >I suspect they would, to stop you messing with their system and causing false positives. Might even throw you in jail on some charge.

            If such systems are even running on the massive CCTV networks yet the volume of false positives just from the volume of people being monitored means I doubt they would even notice – the human supivisor takes that one glance at the CCTV and the guy the AI says it is and goes ‘stupid machine how can you get it so wrong’ and probably doesn’t even notice the shirt is very deliberate and what fooled the machine. And for now I really can’t see such a thing being widespread at all – the energy consumption and computing power required would be far far too costly.

            Also rather hard to throw somebody in jail for wearing clothing, at least in the ‘civilised’ world. The Lawyers would have a field day…

    1. Given the images in this article, seems to me that just wearing the mask, regardless of what pattern is on your clothing, would be sufficient to prevent an AI from recognizing that you have a face, no?

  1. Wiliam Gibson would be proud (cf. “Zero History”), but the real challenge is to get everyone to wear surveillance-defeating patterns, otherwise the wearer would be “that person with the ugly shirt”.

    Also the design has to be non-uniform so that the pattern can’t be recognized and masked out, and also not so random that it provides individuating information (“that person with the ugly shirt with the blob on the upper left shoulder but no stripe in the middle”). Otherwise you’re just wearing your own UPC code unless you get a new one every time you go out. That’s a job for Deep Dream and digital print, not knitting needles.

    One time pad shirt, anyone?

  2. I mean she already demonstrated that it is useless against more modern face detection, over the dumb algorithms that just try to isolate two eyes and a mouth, which is exactly what the design is.

    Realistically there’s no solution to trick future killbot/murderbot/surveillance systems; if we can see a person in an image then a machine certainly can. Not to mention a machine will very likely be capable of doing things we can’t: seeing in all wavelengths including thermal, zooming in with optics that our eyes just don’t have, being able to recount every detail of a scene that it has ever seen to notice any discrepancies, analysing odours and the other countless compounds that our human bodies throw out into the air.

    1. >Realistically there’s no solution to trick future…

      Oh there are, at least for quite some time to come still – as the adversarial images that make the machine ID a lightly modified picture of a tortoise as a rifle etc show. It really needs to get much much smarter first.

      More data from sensors a human won’t have doesn’t make it better at identifying the subject either – more data can be useful but it also gives more attack surfaces for those deliberately adversarial tactics – the machine won’t know with conflicting data which conflicting data is correct. So add a little mylar strip/spots inside your clothes for instance, it will really bugger with the thermal cameras, all those spots that your body heat doesn’t get through, but be entirely invisible to visible light.

      Though as somebody above posted if you are the only person it can’t ID that is as good as an ID in itself – you need to get many folks into the adversarial gear.

Leave a Reply to daveCancel reply

Please be kind and respectful to help make the comments section excellent. (Comment Policy)

This site uses Akismet to reduce spam. Learn how your comment data is processed.