0

Is there a way to change a picture so it would fail to match against existing copy of the same picture on the big social networks such as Facebook, Twitter, Instagram, Tinder and others? Desired result should be as close as possible to the original.

Caballero
  • 103
  • 2

1 Answers1

2

It's dependent on the algorithm used for comparison, but if you have access to that algorithm, you can add small amounts of noise to pixels randomly until it misclassifies the image.

If you had deeper knowledge of the algorithm used, and it was differentiable, you could use gradient descent (ascent) to modify the image to increase the error of what was detected. This is the reverse of what you do when training image comparisons/classifiers which modify the weights to minimize the error.

But in short, if you dont have much knowledge of how the algorithm used works, I personally would try adding noise to pixels randomly, such that the image was still easily recognizable, but a decent amount of noise was added in.

Alan Wolfe
  • 7,711
  • 3
  • 29
  • 72
  • 1
    Probably, there is a fft, wavelet or laplacian. Many of them are extremely noise resistant. – joojaa Nov 17 '19 at 20:23
  • Is there something that those techniques are weak against? I wonder if tone mapping or white balance type things would thwart them? – Alan Wolfe Nov 17 '19 at 20:45
  • Well usually a mirror image or gafting two pictures into one – joojaa Nov 17 '19 at 21:01
  • Ah neat, so maybe mirror image and add some noise to cover your bases? – Alan Wolfe Nov 17 '19 at 21:02
  • 1
    @joojaa The laplacian is the opposite of noise resistant, it blows up noise disproportionately. That's why you usually apply Gaussian convolution before that. But facebook certainly uses something more robust, so it should deal with noise quite easily. Honestly, without knowing the algo, or extensive trial and error, it is very hard to know how to modify the image. – lightxbulb Nov 17 '19 at 21:49