Skip to main navigation Skip to search Skip to main content

MCKD: A Modality Compression Knowledge Distillation Model for Efficient Beam Prediction in V2X Communications

  • Northwestern Polytechnical University Xian

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

Abstract

Beam prediction plays a pivotal role in enabling reliable and low latency communication for Vehicle-to-Everything systems. It typically requires high capacity models trained on multimodal data to achieve accurate and robust performance. However, roadside units face strict computational limits and often have access to only a single sensing modality. These constraints hinder the deployment of heavy multimodal fusion models. To address these challenges, we propose MCKD, a modality compression knowledge distillation model that transfers multimodal representations to a single student model for efficient beam prediction. Specifically, MCKD maintains performance comparable to multimodal models while significantly reducing parameter complexity. Extensive experiments across various weather conditions demonstrate that MCKD achieves high accuracy and adapts robustly to environmental variations.

Original languageEnglish
Title of host publicationICC 2026 - IEEE International Conference on Communications, Proceedings
PublisherInstitute of Electrical and Electronics Engineers Inc.
ISBN (Electronic)9798319542090
DOIs
StatePublished - 2026
Event2026 IEEE International Conference on Communications, ICC 2026 - Glasgow, United Kingdom
Duration: 24 May 2026 → 28 May 2026

Publication series

NameIEEE International Conference on Communications
ISSN (Print)1550-3607

Conference

Conference2026 IEEE International Conference on Communications, ICC 2026
Country/TerritoryUnited Kingdom
CityGlasgow
Period24/05/26 → 28/05/26

Keywords

  • Beam Prediction
  • Generative AI
  • Knowledge Distillation
  • V2X Communications

Fingerprint

Dive into the research topics of 'MCKD: A Modality Compression Knowledge Distillation Model for Efficient Beam Prediction in V2X Communications'. Together they form a unique fingerprint.

Cite this