This is the implementation of the paper SquirRL: Automating Attack Discovery on Blockchain Incentive Mechanisms with Deep Reinforcement Learning (https://arxiv.org/abs/1912.01798).

We test 3 different chain rules: Bitcoin(longest chain rule), Ethereum(uncle block rule) and GHOST (heaviest sub-tree rule).

In \SingleAgent, run

python btc.py
python eth.py
python ghost.py

and you can train the RL agents and test them.

\SingleAgent\environment.py

The interactive environment for 3 different protocols. OpenAI gym type APIs.

\SingleAgent\btc.py , \SingleAgent\eth.py, \SingleAgent\ghost.py

The training and testing part of the RL agent for the Bitcoin/Ethereum/GHOST protocols. You can find more detailed comments of training in \SingleAgent\btc.py.

\SingleAgent\optimal_policy.py

The implementation of Optimal Selfish Mining. We saved the policy in optimal_policy.txt.

We reproduce the two-agent block withholding result in The Miner's Dilemma. Given the proportion of mining power two mining pools possess, there exists a unique Nash equilibirium which they fall in when deciding how much mining power to infiltrate in each other's pool. We use the SquirRL framework to find such equilibrium.

In \BlockWithholding, run

python blockwithholding.py > result.txt

to train the RL agents and find the strategies and rewards of them.

You can specify the mining power of the two agents via argument alphas, e.g., for two mining pools possessing 30% and 50% mining power respectively, you can run the program with --alphas 0.3 0.5.

The policies and rewards of our training result is saved in \result\result.csv.

Ordered by alphabet:

Charlie Hou charlieh@andrew.cmu.edu

Mingxun Zhou : zhoumingxun@pku.edu.cn

Yan Ji : yj348@cornell.edu