This project simulates and evaluates the vulnerabilities of collaborative Multi-Armed Bandit (MAB) algorithms using real-world data from the MovieLens 20M dataset.
It compares a Single-Agent baseline against a Cooperative Multi-Agent network. To rigorously test the network's robustness against data poisoning, the total network size is kept strictly constant (
The framework supports multiple algorithmic policies (UCB, Epsilon-Greedy, and Thompson-Sampling) and utilizes multiprocessing to run multiple attack scenarios concurrently.
To run the simulation, you can execute the Python file directly or use the provided batch/shell scripts (run.bat or run.sh).
By default, the simulation maintains a strict total of 25 agents and tests malicious scenarios at 0%, 20%, 40%, 60%, and 80% network toxicity.
| Argument | Short | Description | Default |
|---|---|---|---|
--data |
Path to the MovieLens ratings.csv file |
data/ml-20m/ratings.csv |
|
--k_arms |
-k |
Number of movies (arms) to process | 1000 |
--t_horizon |
-t |
Total time steps (Auto-calculates to |
Auto |
--policy |
Bandit policy to use (UCB, Epsilon-Greedy, Thompson-Sampling) |
UCB |
|
--explore_param |
Exploration parameter (Alpha for UCB, Decay window for Epsilon) | Auto | |
--beta |
Phase duration scaling parameter | 2.0 |
|
--no_show |
Saves the plots in the background without opening a window | False |
Examples:
- Run Thompson Sampling on a massive dataset in the background:
python main_constant_total.py -k 100000 -t 3500000 --policy Thompson-Sampling --no_show
(Note: Data and plots are automatically saved to the results/ folder with unique filenames like mab_results_ConstantTotal_UCB_K8546_T3500000.npz so previous runs are never overwritten).
If you want to redraw your graphs, isolate specific lines, or magnify the convergence gaps using the intuitive zoom feature.
| Argument | Description |
|---|---|
--file |
The specific .npz file to load from the results/ folder |
--single |
Flag to plot the Single-Agent Baseline |
--coop |
Flag to plot the Fully Cooperative Multi-Agent (0% Malicious) |
--malicious |
Flag to plot the Malicious Scenarios (20%, 40%, 60%, 80%) |
--zoom |
Intuitively zoom in on the end of the timeline (e.g., 80 fast-forwards past the first 80% of the graph and auto-scales the Y-axis to magnify the final 20%) |
--no_show |
Save the newly generated plots without popping up the window |
Example:
- Magnify the Malicious Damage (Exclude Single-Agent & Zoom in 80%):
python utils/replot_constant.py --file mab_results_ConstantTotal_UCB_K8546_T3500000.npz --coop --malicious --zoom 80
To definitively test which mathematical framework (UCB, Epsilon-Greedy, or Thompson Sampling) is the most robust, you can plot them against each other for a specific scenario.
(Note: You must first run main_constant_total.py for all three policies using the exact same parameters so the script has the data to compare!)
| Argument | Description |
|---|---|
--suffix |
The shared end of the .npz filenames (e.g., K8546_T3500000) |
--target |
The specific scenario to compare: single, coop, or the exact 20 for 80% Malicious) |
--zoom |
Zoom in on the end of the timeline (e.g., 80 auto-scales to the final 20%) |
--no_show |
Save the newly generated plots without popping up the window |
Examples:
- Compare the 3 algorithms in a Perfect Cooperative Network:
python utils/compare_policies_constant.py --suffix K8546_T3500000 --target coop
- The Ultimate Stress Test (Compare under a Severe 80% Attack):
python utils/compare_policies_constant.py --suffix K8546_T3500000 --target 20 --zoom 80
To convert the high-speed .npz binary data into a cleanly labeled .csv file (for viewing in Excel or other statistical software):
python utils/export_csv_constant.py --file mab_results_ConstantTotal_UCB_K8546_T3500000.npzThis will generate a matching .csv file right next to it in the results/ folder.

