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

Federated embodied agent learning protects the data privacy of individualvisual environments by keeping data locally at each client (the individualenvironment) during training. However, since the local data is inaccessible tothe server under federated learning, attackers may easily poison the trainingdata of the local client to build a backdoor in the agent without notice.Deploying such an agent raises the risk of potential harm to humans, as theattackers may easily navigate and control the agent as they wish via thebackdoor. Towards Byzantine-robust federated embodied agent learning, in thispaper, we study the attack and defense for the task of vision-and-languagenavigation (VLN), where the agent is required to follow natural languageinstructions to navigate indoor environments. First, we introduce a simple buteffective attack strategy, Navigation as Wish (NAW), in which the maliciousclient manipulates local trajectory data to implant a backdoor into the globalmodel. Results on two VLN datasets (R2R and RxR) show that NAW can easilynavigate the deployed VLN agent regardless of the language instruction, withoutaffecting its performance on normal test sets. Then, we propose a newPrompt-Based Aggregation (PBA) to defend against the NAW attack in federatedVLN, which provides the server with a ''prompt'' of the vision-and-languagealignment variance between the benign and malicious clients so that they can bedistinguished during training. We validate the effectiveness of the PBA methodon protecting the global model from the NAW attack, which outperforms otherstate-of-the-art defense methods by a large margin in the defense metrics onR2R and RxR.

Navigation as Attackers Wish? Towards Building Robust Embodied Agents under Federated Learning