8

If the output of an algorithm when interacting with the [encryption] protocol matches that of a simulator given some inputs, it ‘need not know’ anything more than those inputs.

  • Can a machine learn to find a method to break encryption protocols?
  • How is that possible, to decrypt ciphers using deep learning methods?
R1w
  • 1,952
  • 4
  • 20
  • 45

1 Answers1

16

There is no evidence of deep learning breaking modern cryptography. Deep learning is simply glorified gradient descent. With a reasonable cipher you get no indication of almost finding the key, so I see no hope of deep learning breaking a black box cipher.

In order to use deep learning for cryptography we would need to find a notion of gradually or partially solving the problem, not an easy task.

We have seen some work to use deep learning to build new ciphers, using adversarial training. But there is not much evidence these ciphers are secure against human cryptanalysis.

Ilmari Karonen
  • 46,120
  • 5
  • 105
  • 181
Meir Maor
  • 11,835
  • 1
  • 23
  • 54
  • Did not Alan Turing decrypt the Enigma machine with such an idea? – R1w Sep 08 '18 at 06:33
  • 2
    Enigma had 23.5 bits of key entropy. Enigma is security through obscurity. Enigma can't stand against today's standards of security. And I still don't think deep learning can crack Enigma. – John Dvorak Sep 08 '18 at 06:38
  • @JohnDvorak The idea of creating a ‘thinking’ machine is at least as old as modern computing, if not even older. Alan Turing in his seminal paper ‘Computing Machinery and Intelligence’ laid out several criteria to asses whether a machine could be said be intelligent, which has since become known as the ‘Turing test’. For some great explorations on variants of the Turing test, check-out Brain Christian’s book detailing his adventures with the Loebner Prize entitled The Most Human or check the amazing, dramatized version in Ex-machina.[https://bit.ly/2O0ZTBZ] – R1w Sep 08 '18 at 07:13
  • 4
    @John Dvorak: How do you come to the conclusion that "Enigma had 23.5 bit of key entropy"? Based on this answer, the M3 with 10 wires on the plugboard had 67.1 bit of key entropy (not counting 14.1 bit for the initial, nominally random rotor position, much like we do not count a random IV in the key entropy of a modern cipher). That source gets 87.5 bit for the M4 with 13 wires. – fgrieu Sep 08 '18 at 09:04
  • 1
    Ah... somehow I thought all that was configurable on Enigma was the rotation of five wheels of letters. Still, 70 bits of entropy wouldn't be sufficient in a today's cypher. – John Dvorak Sep 08 '18 at 09:17
  • 1
    If you think that deep learning is simply running down a hill, watch AlphaGo thrash the world's best Go player. Or a learning network invent a whole new way to win at Breakout. Crypto analysis can't be far away. It's a fairly simple and closed domain after all. – Paul Uszak Sep 08 '18 at 21:39
  • 4
    @PaulUszak You are comparing apples and oranges. With Go, you can tell when you are winning or losing, and each move gives you feedback. Do you still think deep learning could excel at Go if it wasn't allowed to see the opponent's response or the current board state, and had to win a trillion trillion trillion trillion games in a row before it is able to learn that it has even one a single game? – forest Sep 09 '18 at 02:56
  • 1
    @forest But it can see the output. – Paul Uszak Sep 09 '18 at 16:04
  • 1
    @forest And in the whole domain of the sciences, apples and oranges are virtually identical aren't they? – Paul Uszak Sep 09 '18 at 16:06
  • 4
    Seeing the output isn't likely to produce a meaningfull gradient. It's not that this hasn't been tried. It doesn't work. – Meir Maor Sep 09 '18 at 17:35
  • 2
    @PaulUszak Seeing the state of a Go tournament means you are able to see, at all times, how well you are doing. It's like a cipher that tells you how close you are to breaking it. For a real cipher, you don't see the result until you have successfully performed some form of cryptanalysis. Comparing a game that is constantly revealing its state with a complex chaotic mathematical system that is designed specifically to give away no information is just silly. The fact is, neural networks can only attack the most basic problems, like that of Go and Chess. – forest Sep 10 '18 at 01:37
  • 3
    @JohnDvorak There are still unsolved messages today (from WW2) for Enigma and people have tried to break them with no success.... You still need some luck to break them or long enough messages. So in some cases it's enough. – Henno Brandsma Sep 13 '18 at 21:57