English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Zendy Plus

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Zendy Tools

Zendy Open

Most existing works focus on direct perturbations to the victim'sstate/action or the underlying transition dynamics to demonstrate thevulnerability of reinforcement learning agents to adversarial attacks. However,such direct manipulations may not be always realizable. In this paper, weconsider a multi-agent setting where a well-trained victim agent $\nu$ isexploited by an attacker controlling another agent $\alpha$ with an\textit{adversarial policy}. Previous models do not account for the possibilitythat the attacker may only have partial control over $\alpha$ or that theattack may produce easily detectable "abnormal" behaviors. Furthermore, thereis a lack of provably efficient defenses against these adversarial policies. Toaddress these limitations, we introduce a generalized attack framework that hasthe flexibility to model to what extent the adversary is able to control theagent, and allows the attacker to regulate the state distribution shift andproduce stealthier adversarial policies. Moreover, we offer a provablyefficient defense with polynomial convergence to the most robust victim policythrough adversarial training with timescale separation. This stands in sharpcontrast to supervised learning, where adversarial training typically providesonly \textit{empirical} defenses. Using the Robosumo competition experiments,we show that our generalized attack formulation results in much stealthieradversarial policies when maintaining the same winning rate as baselines.Additionally, our adversarial training approach yields stable learning dynamicsand less exploitable victim policies.

Rethinking Adversarial Policies: A Generalized Attack Formulation and Provable Defense in RL