ban img

I'm currently playing with GANs. GANs are awesome, and the shortcoming GANs have shown previously are slowly getting fixed thanks to amazing researchers. GANs, if you don't know already, are a way to have a neural net to (hopefully) create realistic content.

It's no secret now that I'm working for an adult company, and that gives a quite funny color to the research I'm doing, As such, well, I thought "dude, let's use GANs to try and generate adult content. Latent space interpolation will be funny as heck.". So that was it. I reused a part of the dataset I'm using for action / position recognition, and threw a DCGAN on it with the newly proposed R1 regularizer.

A day later, I thought the images were pretty unsettling and weird, while being faaaaaar away from any recognizable adult content. Still very abstract to the human eye. So, well, let's post it on facebook, right? That'll amuse my friends while still being SFW.

And the second I posted it, Facebook's anti-porn filter banned me for three days. The key part of the previous paragraph was to the human eye. What I naively didn't expect was that it was actually really close to porn in convnet feature space, and I got banned for seemingly abstract pictures.

So there you have it folks: a good ol' black box adversarial attack to Facebook's anti-porn filter. Well, let's rather call it an adversarial suicide.

Next step is to actually attack the system and be able to post porn without being flagged. Just for science. Just to make a point about how unsafe neural nets currently are.