Google’s AI Can Now Invent Its Own Cryptography


Dr. Alfredo Carpineti

Senior Staff Writer & Space Correspondent

clockOct 31 2016, 20:15 UTC


Two artificial intelligences (AIs) have developed a way to talk to each other in secret, stopping a third AI from spying on them. Although this might sound like the plot of a sci-fi novel, it’s actually a real breakthrough Google engineers have achieved with their smart machines.

Researchers Martín Abadi and David Andersen from Google Brain asked two artificial neural networks, Alice and Bob, to use secret keys to share private information among themselves while an adversary network named Eve tried to intercept the communication.


The setup had messages from Alice getting to both Bob and Eve. For the first 7,000 messages, Eve was able to grasp as much of the message as Bob, but afterward, Alice and Bob learned how to talk privately more quickly than Eve was capable of eavesdropping. In the long run, Bob could understand every single character sent by Alice, while Eve was only getting slightly more than half of them.

content-1477932765-1477754298-bob-eve-goThe ability of Bob to correctly guess Alice's messages increased quickly over time, while Eve didn't improve much. Abadi & Andersen/Google

In a paper, available online, the researchers explain that the networks weren’t given a specific set of cryptography algorithms to try. They simply devised a way to keep their communication secret based exclusively on the training objectives put forward by the researchers: stopping Eve.


Neural networks continue to astound with their unique approach to problem solving. The architecture of these basic AIs mimics the neuron distribution in animal brains, and computer scientists have shown that this approach creates machines with an incredible potential for self-learning.

The neural networks are trained in specific tasks and some of them become remarkably good at what they do. An example of this is Google’s AlphaGo, which became a world champion in the ancient game of Go earlier this year.

The researchers think these neural networks should be studied in detail. They might not be great at cracking codes, but their unusual strategy in creating them might be very useful for protecting information.

  • tag
  • artificial intelligence,

  • machine learning,

  • AlphaGo,

  • spying,

  • neural network,

  • smart machines,

  • eavesdropping