2 ms·
It's funny how a very simple DL architecture any beginner could prepare ended up in a prohibition act - essentially training two convolutional autoencoders, one
by bitL 8y ago
It's funny how a very simple DL architecture any beginner could prepare ended up in a prohibition act - essentially training two convolutional autoencoders, one on the face one wants to replace (extracting dataset by e.g. tracking face with OpenCV in target video), the second on many pictures of the face one wants to replace (e.g. scraping from Internet or known videos), then essentially gluing together encoder from one and decoder from the other, assuming latent variables match, and suddenly a fake video is done.
IMO if you really want to control deep fakes, you should use a blockchain and track all processing steps, starting from image/video acquisition. I don't want to give anyone ideas, but they will do it anyway.
- dieblur 8y agoThis doesn't even begin to make sense. You can't force people to register any creation on the blockchain and there can't be a main "only real things allowed" chain. If we want to shoehorn a blockchain into every possible problem ever, lets begin with ones that could actually work.
- bitL 8y agoThe idea is to enforce equipment manufacturers to prepare uniquely identifiable IDs in all imagery/footage their equipment produces, entice all software vendors to register all operations in a blockchain and store a reference to it in metadata; then reject any footage that can't have all steps verified as "fake". How is that any different than tracking production inputs in logistic chains using blockchain?
- xamuel 8y agoNothing can stop me from pointing a camera at my computer monitor and recording a verified video of a non-verified video.
- Skrillex 8y agoI believe that the usual response to this is that the devices include some kind of unmodifiable time and location stamping, although that argument spirals out of control pretty quickly.
- bitL 8y agoSure, but in that case your chain starts with your camera, producing "pro" looking video which would be totally unlikely (not mentioning trivial stuff like noticing monitor/lens distortions etc.). Analyzing subsequent processing steps stored in blockchain would likely show very low probability of authentic work due to missing many required steps (OK, there is still certain low probability you can fool it somehow, but would you want to waste so much time/processing power on it?)
- xamuel 8y agoThis is whack-a-mole. If I can design deepfake software, it isn't much harder to design it to specifically anticipate the user filming the result. The user would input their monitor specs, their camera specs, etc., and the software would produce a weirdly distorted video which looks perfect when filmed with that camera from that monitor. Or, people just outright sell doctored cameras where you can intercept the input feed. This isn't like adult content filtering, where all that matters is that kids can't get around it. You have to assume people who know what they're doing are going to attack your technical solution, and apply ingenuity in doing so. When the enemy consists of hackers, you can't ward them off with a hack!