yeldarb

28 ms·

241.
by yeldarb 5y ago
We had pretty good success on https://paint.wtf (AI-judged Pictionary game) using CLIP for content moderation. Feel free to probe it a bit by drawing a dick or something on one of the prompts; it seems to do a pretty good job (i
242.
by yeldarb 5y ago
Hey, OP here - while that is also true, I think the average observer would think this number _should_ be zero. The fact that any exist at all is surprising (but the fact that only two exist is also surprising; I expected to find either zero
243.
by yeldarb 5y ago
It doesn't seem like it; check out the adversarially constructed images here. They don't look anything like the original despite perfectly matching the NeuralHash: https://github.com/AsuharietYgvar/AppleNeural
244.
by yeldarb 5y ago
The instructions are here: https://github.com/AsuharietYgvar/AppleNeuralHash2ONNX This was discussed on HN yesterday here (and a few other front-page stories): https://news.ycombinator.com/item?id=28218
245.
by yeldarb 5y ago
The sample set is ImageNet, which is a well-known dataset in Computer Vision and is available for download here: https://www.kaggle.com/c/imagenet-object-localization-challe... I'd love to see this work extended;
246.
by yeldarb 5y ago
Their reported collision rate was against their CSAM hash database[1]. > In Apple’s tests against 100 million non-CSAM images, it encountered 3 false positives when compared against NCMEC’s database. In a separate test of 500,000 adult p
247.
by yeldarb 5y ago
Different things; birthday paradox was testing them against each other, turning it into a sorting problem was the other.
248.
ImageNet contains naturally occurring Apple NeuralHash collisions(blog.roboflow.com)
860 points by yeldarb 5y ago|510 comments
249.
by yeldarb 5y ago
Those are examples of triggering false negatives (CSAM that gets missed by the system), this collision attack is about triggering false positives (non-CSAM that triggers human intervention).
250.
by yeldarb 5y ago
Doing that would mean the NeuralHash would change though. And you'd have to not only get CLIP to identify CSAM in the generated image but also negate the parts of the generated image that are causing CLIP to label it as "generated
251.
by yeldarb 5y ago
I created a proof of concept showing how OpenAI's CLIP model can function as a "sanity check" similar to how Apple says their server-side model works. In order for a collision to get through to the human checkers, the same im
252.
by yeldarb 5y ago
I had a lot of fun experimenting with VQGAN+CLIP a couple of weeks ago too; it feels kind of like you're prodding into the mind of an alien species to extract its ideas. My findings: https://blog.roboflow.com/ai-generat
253.
by yeldarb 5y ago
I was curious how big of a deal this would be in a real-world attack (eg could someone use this to DDoS Apple's human reviewers?) so I ran the colliding images through another feature extraction model, OpenAI's CLIP, to see if the
254.
Mitigating the Collision of Apple's CSAM NeuralHash(blog.roboflow.com)
5 points by yeldarb 5y ago|0 comments
255.
by yeldarb 5y ago
I recently wrote up some of my experiments, observations, and findings with VQGAN+CLIP (the models folks have been using to do this AI Art recently) here: https://blog.roboflow.com/ai-generated-art/
256.
by yeldarb 5y ago
What are you supposed to _do_ with it? Is it for water-skiing/wakeboarding? Tubing? Just cruising? What market segment is it trying to play in? I don't think I've ever really seen a 24 foot closed-bow boat.
257.
Experimenting with CLIP+VQGAN to create AI generated art(blog.roboflow.com)
2 points by yeldarb 5y ago|0 comments
258.
by yeldarb 5y ago
This was our (high school) intern, Samrat Sahoo's, summer project at Roboflow. He collected and labeled a dataset of rabbits in his backyard, then trained a computer vision model to identify them and wired up a Raspberry Pi to the mode
259.
Show HN: Protecting My Garden from Rabbits with Computer Vision(blog.roboflow.com)
11 points by yeldarb 5y ago|2 comments
260.
by yeldarb 5y ago
I believe the current understanding of GAN copyright is that the "minimum degree of creativity" happens when a human chooses the inputs/outputs and copyright is assigned to the human at that point. Drawing the input image for
261.
by yeldarb 5y ago
An example is here (using the same NVIDIA GauGAN model that backs this Canvas app): https://twitter.com/jonathanfly/status/1144735290591981568?l... Jonathan has played with GauGAN quite a bit (search twitter for &
262.
by yeldarb 5y ago
And then once you have the weights you have to figure out what to do with them. Data science (training a good model) and engineering (deploying it as a service to be consumed by end users) are pretty distinct tasks.
263.
by yeldarb 5y ago
I’ll be excited for the day when support for worker threads is more widespread. Tried to use them with Node Canvas a few months ago but it’s not context aware[1] so had to go back to forking processes. Intriguing to see someone in this thre
264.
by yeldarb 5y ago
Is there something equivalent to Nomadlist but for filtering places to live in? I don't know anything about Youngstown or Jackson but based on Flint's reputation I imagine there's probably a reason it's a cheap place to
265.
Stable Rings from Rectangular Lego Bricks (2001)(people.idsia.ch)
2 points by yeldarb 6y ago|0 comments
266.
by yeldarb 6y ago
We tried out the demo last summer (which was a bit of manu-mation before the product was fully built out -- kudos to Lenny & crew for doing things that don't scale) and had a great experience! Here's an example of one of the v
267.
by yeldarb 6y ago
Nice resource, I submitted a PR with our free-tier offerings!
268.
by yeldarb 6y ago
So many people are eager to steer a conversation. That’s not Aaron. He’s one of the best listeners I have ever met. And he’s one of those people that engenders so much respect that it makes you want to be better just because you don’t want
269.
by yeldarb 6y ago
It sure is! Roboflow builds tools that enable any developer to use computer vision without having to become a machine learning expert. (Kind of like how you don't have to become a cryptography expert to add auth & accept payments i
270.
by yeldarb 6y ago
Were there any gaps in the guesses? You have to be first to guess the correct answer so taking an un-picked lottery number in the vicinity of the median would be a decent strategy.

More ›