Shielded RL: A Safe Reinforcement Learning Concept

Shielded RL: A Safe Reinforcement Learning Concept

Presenter(s)

Shruti Singh

Comments

12:40-1:00, LTC Studio

Description

Artificial Intelligence (AI) continues to be a part of our everyday life in ways one thoughtimpossible. From instantly curated music playlists and virtual experiences to autonomousvehicles on the roads, everyday life has become intertwined with AI’s remarkablecapabilities. A significant driving force behind these advancements is reinforcement learning(RL), a field of AI that excels at conquering challenging tasks through interactions withdynamic environments. Its adaptability has inspired solutions—from steering driverless carsand maneuvering robot fleets to managing warehouse logistics and guiding patienttreatments—reinforcement learning effortlessly demonstrates its power across diversedomains. Unfortunately, similar to any popular field, RL-based systems also draw theattention of malicious actors who look for cracks in these sophisticated models. Inhigh-stakes situations, malicious noise injected at precisely the right time can compromisethe integrity of an RL system, leading to potentially disastrous outcomes. Think of aself-driving car receiving deceptive signals about obstacles or lane markings—turning whatshould be a smooth journey into a hazard for passengers and pedestrians alike. This kind ofstrategic tampering, known as an adversarial attack, can derail an RL agent if we fail to usetailored defense mechanisms. Although there is no universal, foolproof remedy foradversarial onslaughts, developing strategies that detect, evaluate, and counter suchthreats—based on the specific needs and vulnerabilities of a given system—forms thebedrock of robust protection. This brings us to the idea of shielded reinforcement learning(shielded RL), which acts like a security gate that stands between the learning agent andpotential disaster. Much like the way we described adversarial attacks in everyday terms,shielded RL can be understood as a “safety net” for the agent: it continually monitorspossible actions and prevents reckless or unsafe choices before they become catastrophic.In other words, while the RL agent learns by trial and error, the shield remainsever-vigilant—subtly guiding it around pitfalls and ensuring that it does not stray intodangerous territory. By embracing shielded RL, we can build robust reinforcement learningsystems that not only perform impressively, but also stay safe against determined attackers.

Publication Date

4-23-2025

Project Designation

Graduate Research

Primary Advisor

Luan V. Nguyen, Tam Nguyen

Primary Advisor's Department

Computer Science

Keywords

Stander Symposium, College of Arts and Sciences

Institutional Learning Goals

Community; Diversity; Practical Wisdom

Shielded RL: A Safe Reinforcement Learning Concept

Share

COinS