RAMPART-FL: A Dataset for Offline Reinforcement Learning in Federated Participant Selection

Dataset Summary

This repository contains a dataset package generated by the RAMPART-FL framework, a system for researching Reinforcement Learning (RL) based participant selection in Federated Learning (FL) for intrusion detection. The data originates from a 400-round, 25-client simulation where an RL agent used a Multi-Criteria strategy to select clients.

This package provides two distinct files to serve different research needs:

1. rampart_fl_400r_25c_multicriteria_event_log.csv: A detailed, event-driven log that is ideal for in-depth analysis, custom feature engineering, and understanding the step-by-step behavior of the FL process.
2. rampart_fl_400r_25c_multicriteria_rl_transitions.csv: A pre-processed, analysis-ready dataset of (State, Action, Reward, Next_State) tuples that has been enriched with contextual client and reward information.

The scripts used to generate these files from the raw experimental output are available in the associated GitHub repository.

Associated Research

This dataset is a key contribution of a master's thesis focused on creating robust and adaptive security solutions for IoT networks using Federated Learning.

• Framework: RAMPART-FL on GitHub

---

File 1: The Refined Event Log (rampart_fl_400r_25c_multicriteria_event_log.csv)

This file provides a detailed log of all relevant events that occurred during the simulation. It is best suited for researchers who need to perform detailed analysis beyond standard RL transition data.

The dataset is structured in a sparse, event-driven format. This means that for any given row, only the columns relevant to its event_type will contain data. All other columns in that row will be empty.

Event Log Columns

Column Name	Description	Data Type
Core Identifiers
elapsed_seconds_since_start	Time in seconds since the simulation began.	Float
server_round	The FL training round number.	Integer
event_type	The type of event being logged (selection_info, learning_update, etc.).	String
client_cid	The unique identifier for the client.	Integer (Large)
State, Action & Policy (selection_info)
s_client_state_tuple	The state vector (S) representing the client's condition.	String
s_was_selected	The action (A) taken by the agent (1 or 0).	Integer
s_available_cids_count	The number of clients available for selection.	Integer
s_client_q_value	The Q-value of the state-action pair from the original agent.	Float
s_client_selection_prob	The selection probability from the original agent's policy.	Float
Reward & Learning (learning_update)
l_state_at_selection	The state (S) for which the reward is being applied.	String
l_reward_for_action	The total global reward (R) for the round's actions.	Float
l_reward_*_component	The different components (performance, fairness, etc.) of the total reward.	Float
Client Evaluation Metrics (client_eval_metrics)
c_eval_partition_id	The client's static data partition ID.	Integer
c_eval_profile_name	The client's hardware profile name.	String
c_eval_cores	Number of CPU cores available to the client.	Integer
c_eval_f1, _accuracy, etc.	The client's local evaluation performance metrics.	Float
c_eval_num_samples	Number of samples in the client's local test set.	Integer
Client Fit Metrics (client_fit_metrics)
c_fit_time_seconds	Time taken for the client's local training task.	Float
c_fit_cpu_percent	The client's average CPU usage during training.	Float

---

File 2: Enriched RL Transitions Dataset (rampart_fl_400r_25c_multicriteria_rl_transitions.csv)

This file provides a clean, analysis-ready dataset where each row is a complete, enriched (State, Action, Reward, Next_State) tuple. It is designed for direct use in most offline RL training libraries and workflows. Contextual columns (like c_eval_f1) will only contain data if the action for that row was 1.

Transitions Dataset Columns

Column	Data Type	Description
Core Transition
server_round	Integer	The server round N in which the state was observed and the action was taken.
client_cid	String	The unique identifier for the client.
state	String	The state vector (S_t) representing the client's condition at the start of round N.
action	Integer	The action (A_t) taken for the client: 1 if selected, 0 otherwise.
reward	Float	The global reward (R_{t+1}) received after the completion of round N.
next_state	String	The subsequent state vector (S_{t+1}) for that same client at the start of round N+1.
Reward Components
l_reward_performance_component	Float	The portion of the reward attributed to the global model's performance.
l_reward_fairness_penalty_component	Float	The portion of the reward attributed to the fairness penalty.
l_reward_resource_cost_component	Float	The portion of the reward attributed to the resource cost of selected clients.
Client Context (for the round)
c_eval_profile_name	String	The client's hardware profile name (e.g., High-End Edge CPU).
c_eval_cores	Float	Number of CPU cores available to the client (Float due to possible NaNs).
c_eval_f1	Float	The client's local F1-score from its evaluation in this round.
c_eval_num_samples	Float	Number of samples in the client's local test set.
c_fit_time_seconds	Float	Time taken for the client's local training in this round.
c_fit_cpu_percent	Float	The client's average CPU usage during training.