Alice and Bob in Cipherspace
A new form of encryption allows you to compute with data you cannot read
Gentry described his FHE system in his doctoral dissertation and in a paper at the Symposium on the Theory of Computing in 2009. In the three years since then, dozens of variations, elaborations and alternative schemes have been published, along with at least three attempts to implement homomorphic encryption in a working computer program.
Most of the systems share the same overall architecture, with a somewhat homomorphic scheme that gets promoted to full homomorphism. Where the ideas differ is in the underlying cryptographic mechanism—the way that bits are twiddled and secrecy is achieved.
Every cryptosystem is based on a problem that’s believed to be hard in general (so that Eve can’t solve it) but easy if you know a shortcut (so that Alice and Bob can decrypt messages efficiently). RSA’s hard problem is the factoring of large integers; the shortcut is knowledge of the factors. Gentry’s 2009 algorithm relies on a problem from the theory of integer lattices—sets of discrete points arranged like the atoms of a crystal in a high-dimensional space. Lattices give rise to an abundance of computationally difficult problems. For example, from a random position in space it is hard to find the closest lattice point unless you happen to know a specific set of coordinates that serve as a geometric guidebook to the lattice.
In 2010 another homomorphic cryptosystem was invented by Marten van Dijk of MIT, Gentry, Shai Halevi of IBM and Vinod Vaikuntanathan, now at the University of Toronto. In this case the hard problem comes from number theory; it’s called approximate GCD. The exact GCD, or greatest common divisor, is easy to calculate; Euclid gave an efficient (and famous) algorithm. A “noisy” version of the problem seems to be much harder. If two large numbers have the GCD p, and you alter those numbers by adding or subtracting small random quantities, it becomes difficult to find p. In the cryptosystem, p is the secret key.
A problem called learning with errors forms the basis of a third FHE system introduced by Zvika Brakerski of Stanford and Vaikuntanathan. Here the task is to solve a system of simultaneous equations where each equation has some small probability of being false. As with GCD, this is an easy problem in the exact case, where there are no errors, but searching for a subset of correct equations is laborious.
More recently, Brakerski, Vaikuntanathan and Gentry have developed a variant of the learning-with-errors system that takes a different approach to noise management. Instead of stopping the computation at intervals to re-encrypt the data, they incrementally adjust parameters of the system after every computational step in a way that prevents the noise level from ever approaching the limit.