I am so excited about this project. Yes, I am a backer. I love that they are actually testing patterns against real-world cameras (Flocks, Rings, etc) and releasing only those that reliably defeat them.
Check the two most recent updates
The GPUs have been at 100% for months. Billions of patterns tested against an “impossible” gauntlet of camera models. Here is the plain truth on where we stand.
Hoodie: our best piece. Across all the person detectors we test, nearly 5,000 patterns hit a perfect record. The hoodie covers the most of you, so the pattern has the most to work with.
T-shirt: strong on individual detectors, still tightening. T-shirt patterns beat person detectors on tens of thousands of tests. What we are still building is the harder win: a single T-shirt pattern that stops a real camera on every person, every time. The hoodie is there. The T-shirt is close, and the ship design will be our strongest.
Buff: oh, the buff. The hard one, and here is why. It has to beat two things the shirt and hoodie do not: face detection and facial recognition, both far tougher than spotting a person. So far it reliably beats one of the four face detectors (2,600+ wins), facial recognition is cracking. Two of the toughest models have not fallen to our main method yet. So earlier this week I implemented a second, smarter method, and it is already working. In testing, it fooled the hardest face detector 71% of the time on people it had never seen before. That is the direction the final buff design is heading.
I do not count a win unless two things are true. One, the pattern beats a plain black version of the same garment, so we know the pattern is doing the work, not just covering you. Two, it works on people the system was never tuned on That is a strict test. I watch the wins and losses come in live, and most patterns die on that screen. The ones that make it are real.
The shirt and hoodie are proven. They defeat real cameras today. The buff is the last wall, and I have found the cracks. Last night, on faces it had never seen, it beat that detector 71% of the time. A week ago that number was zero.
This is what it looks like when a wall comes down. Not a press release months from now. Right here, in real time, with you watching. We will finish this.
We crossed $100k today. I set the goal at $5,000 because that is what I thought it would take to get the first patterns onto real fabric. I never expected to be writing this kind of update.
When I started this research I knew it was important. The feedback from the community has been overwhelming, and your support for this project means you care about your privacy too. Know that every waking minute I am developing the strongest countermeasures possible.
Every dollar of this goes into the research. More cameras to test against, more fabric, more print runs, and the time to keep pushing on the problems that are not solved yet. Six months ago I was begging for GPU hours. Now I can chase the hard cells until they break.
Thank you. I am humbled, grateful, and thankful for you all. Genuinely. If you know someone who would want this, please send it their way.
-Bill
Ummm… no.
IIUC, the vendor would need to retrain their model, not simply apply a patch like for an out-of-bounds memory error.
Sure, but do you remember DeepDream? The whole point of that system was pattern recognition (especially faces) in images. That’s over a decade old.
The point being, image pattern models are legacy machine learning technology. They’re very easy to train compared to the cutting edge language models, you can do it at home with a decent GPU or TPU. Flock and the other companies wouldn’t have to completely dump their existing models either, just train a new one to deal with these patterns specifically and then add it to the system. It really won’t take that much time or resources to address.
Depends.
Do the existing cams allow for over-the-wire model updates? I’d be a bit surprised if that were the case.
Hmm, I kind of assumed the actual recognition model was running on the server side since the cameras are all networked for data sharing (which is kind of the problem).
It’s a fair question, I don’t know enough about their software to speculate.