Fed-2

Fed-2 is a lightweight gossip-based FL system that does not rely on a central server, allowing Fediverse instances to co-train models with other compatible instances in a privacy-preserving manner. Please follow the steps below to run Fed-2.

Step 1: Deploy and Run Your Instance

Please refer to the official documentation to deploy an instance (take Mastodon Instance as example), here are some links and documents you can refer to:

• https://github.com/mastodon/mastodon
• https://github.com/McKael/mastodon-documentation/tree/master/Running-Mastodon
• https://github.com/McKael/mastodon-documentation/blob/master/Running-Mastodon/Development-guide.md
• https://github.com/McKael/mastodon-documentation/blob/master/Running-Mastodon/Production-guide.md

You should now be able to run your Mastodon instance with the following instructions:

• Development env: RAILS_ENV=development bundle exec rails server
• Production env: RAILS_ENV=production bundle exec rails server

Step 2: Install Related Library

python==3.9.15, Flask==2.2.5, joblib==1.2.0, pandas==1.5.2, APSheduler==3.10.4. scikit-learn==1.2.2, sentence-transformers==2.2.2, numpy==1.23.4, requests==2.28.1, pyarrow==8.0.0

Step 3: Setting Instance Information

On lines 20-28 of app.py, you need to specify:

1. Your Mastodon folder path
2. Your instance name
3. Your pre-labeled training data, use to train local model
4. The embedding model, use to transfer post content to embedding

Step 4: Move .rake File

Please move run_Fed2.rake files to /mastodon/lib/tasks/ path. By doing this you can run Fed-2 through the Mastodon CLI.

Step 5: Run Fed-2

First go to the mastodon directory: cd mastodon/

Then run Fed-2 with the following command:

RAILS_ENV=development Fed2:run_Fed2 trainer=YOUR_TRAINER calculator=YOUR_CALCULATOR gossip_peer_num=YOUR_GOSSIP_PEER_NUM gossip_data_time=YOUR_GOSSIP_DATA_TIME calculate_similarity_time=YOUR_CALCULATE_SIMILARITY_TIME federated_partner_num=YOUR_FEDERATED_PARTNER_NUM federated_learning_time=YOUR_FEDERATED_LEARNING_TIME

In the command, you need to set some parameters, of course you can use the default value, the following is the description of each parameter:

• trainer=YOUR_TRAINER
  • trainer is local model training algorithms
  • Optional values are NB (Naïve Bayes) and LR (Logistic Regression), the default value is NB
  • Example: trainer=NB
• calculator=YOUR_CALCULATOR
  • calculator is the metric for calculating instance compatibility
  • Optional values are Hashtags, Peers, Rules and Blocks, the default value is Hashtags
  • Example: calculator=Hashtags
• gossip_peer_num=YOUR_GOSSIP_PEER_NUM
  • gossip_peer_num is the number of gossip peers per instance
  • If 0 < gossip_peer_num < 1, it represents the percentage. If gossip_peer_num > 1, it represents the actual number. The default value is 0.05
  • Example: gossip_peer_num=0.05, it represents gossip to 5% of peers
• gossip_data_time=YOUR_GOSSIP_DATA_TIME
  • gossip_data_time is the time in seconds between each gossip data interval
  • The default value is 3600
  • Example: gossip_data_time=3600
• calculate_similarity_time=YOUR_CALCULATE_SIMILARITY_TIME
  • calculate_similarity_time is the time in seconds between each instance of compatibility evaluation
  • The default value is 7200
  • Example: calculate_similarity_time=7200
• federated_partner_num=YOUR_FEDERATED_PARTNER_NUM
  • federated_partner_num is the number of federated partners that are performing the FL
  • If 0 < federated_partner_num < 1, it represents the percentage. If federated_partner_num > 1, it represents the actual number. The default value is 0.05
  • Example: federated_partner_num=0.05
• federated_learning_time=YOUR_FEDERATED_LEARNING_TIME
  • federated_learning_time is the time in seconds between each FL performed
  • The default value is 10800
  • Example: federated_learning_time=10800