4 ms·
Would better cameras circumvent the issue? Sure a deepfake is convincing on a low resolution compressed video, but what about a 4k or 8k video where lens disto
by colechristensen 6y ago
Would better cameras circumvent the issue?
Sure a deepfake is convincing on a low resolution compressed video, but what about a 4k or 8k video where lens distortion and fine details are everywhere? I don't have a lot of confidence that facial pores could be convincingly simulated.
- SamBam 6y agoBut are most secret camera recordings using high-res cameras? Let's say there was video recording of a politician in a hotel in Russia, to pick a random example. We would already expect it to be low-quality video even if it was real.
- colechristensen 6y ago>Let's say there was video recording of a politician in a hotel in Russia, to pick a random example. We would already expect it to be low-quality video even if it was real. But why? The cheap last-gen iPhone in my pocket can take 4k videos, almost everybody has something like it in their pockets. Quality is cheap and will continue spreading, fuzzy video that looks like it was taken on a 90s camcorder will itself seem suspicious.
- ma2rten 6y agoGenerating those low level feature is actually a much easier problem. Here is an example I found using a google image search for "upscaling": https://copyrightimage.com/2018/05/09/better-image-upscaling-method-in-photoshop/ https://copyrightimage.com/2018/05/09/better-image-upscaling... This is from 2018, which is an eternity for machine learning research.
- colechristensen 6y agoBut those kinds of things also could reveal the fraud - adding a blemish where the real person doesn't have one, features that move around frame by frame in a video. The difference is things which look believable might be details which are trivially verifiably false - sure you can simulate my face and my voice, but can you accurately simulate every pore on my nose. Here's a random example: https://ieeexplore.ieee.org/document/6619284 https://ieeexplore.ieee.org/document/6619284 They can detect a persons pulse from tiny variations in head movement, I'd like to see a deepfake simulate that! And it's not about this one thing, I'm sure you could design against these tiny measurements one by one... the key is that there are a lot of them and a deepfake is going to have to simulate an enormous amount of data in order to pass what will become trivial tests for realism. This is made a much easier weapon against deepfakes when you have progressively better videos.