Reinforcement Learning Attack Demo

Overview

This project is an interactive browser-based demonstration of two reinforcement learning (RL) attack vectors:

• Reward Hacking (via high-reward tiles)
• Adversarial Observation Manipulation (via fake goals and noise)

The demo visualizes how an RL agent’s behavior changes under these conditions.

---

How to Run the Demo

Go to https://sarahhamer.github.io/ConfuseTheLostBot/ and have fun. No extra setup is required.

---

Dependencies

This project has no external dependencies.

It runs entirely in the browser using:

• HTML
• CSS
• Vanilla JavaScript

---

How to Trigger Each Demonstration

1. Reward Hacking (Golden Tiles)

• Increase "Tile Reward" using the slider.
• Optionally click on the grid to place additional gold tiles.

2. Adversarial Observation Manipulation

Fake Goal

• Enable "Fake Goal (Observation Attack)"
• The agent will pursue an incorrect (fake) goal.

Noise

• Increase "Noise"
• The perceived goal becomes unstable or random.

---

Expected Behavior / Output

Normal Behavior

• The agent moves efficiently toward the true goal (green tile) and remains there.

Reward Hacking

• With high tile reward:
  • The agent prioritizes gold tiles over the true goal.
  • It may detour or remain near high-reward tiles.

Fake Goal Attack

• The agent moves toward the blue outlined (fake) goal instead of the true goal.

Noise

• As noise increases:
  • The agent’s path becomes erratic or unstable
  • It may change direction frequently due to incorrect observations.

---

Interaction Notes

• Click on grid cells to add/remove reward tiles
• Use Step to move one step at a time
• Use Run to observe continuous behavior
• Use Reset to restore default state