Evolution

Now that we have a concrete definition for a censorship evasion strategies and a way to use them, we can begin evolving new strategies with Geneva’s genetic algorithm.

A genetic algorithm is a type of machine learning inspired by natural selection. Over the course of many generations, it optimizes a numerical fitness score of individuals; within each generation, individuals that have a low fitness will not survive to the next generation, and those with a high fitness score will survive and propagate.

Each individual in Geneva is a censorship evasion strategy. evolve.py is the main driver for Geneva’s genetic algorithm and maintains the population pool of these strategies.

Each generation is comprised of the following:
  1. Mutation/Crossover - randomly mutating/mating strategies in the pool
  2. Strategy Evaluation - assigning a numerical fitness score to each strategy - decides which strategies should live on to the next generation
  3. Selection - the selection process of which strategies should survive to the next generation

For more detail on mutation/crossover or the selection process, see our papers or the code in evolve.py for more detail. evolve.py exposes options to control hyperparameters of mutation/crossover, as well as other advanced options for rejecting mutations it has seen before.

Strategy evaluation is significantly more complex, and will be covered in depth in the next section.

For most of this documentation, we will show examples of using evolve.py with the --eval-only <strategy_here> option. This instructs evolve.py not to start the full genetic algorithm, but to instead perform a single strategy evaluation with the given parameters. --eval-only can be replaced with parameters to the genetic algorithm to start strategy evolution.

Argument Parsing

Geneva uses a pass-through system of argument parsing. Different parts of the system define which arguments they care about, and they will parse just those args out. If --help is used, evolve.py will collect the help messages for the relevant components (evaluator, plugins, etc).

Population Control

The two most important command line options for controlling evolutions are --population and --generations. These control the population size and number of generations evolution will take place for, respectively.

Geneva will not automatically halt evolution after population convergence occurs.

Seeding Population

Geneva allows you to seed the starting population pool of strategies using the --seed <strategy> parameter. This will create the initial population pool with nothing but copies of the given strategy. Mutation is applied to each individual before evaluation begins, so the first generation is not solely one individual being evaluated over and over again.