Scholarlink

This work introduces a privacy-preserving framework that integrates foundation models with federated learning through a synergistic application using Low-Rank Adaptation (LoRA). Our three-stage pipeline—centralized pretraining, federated fine-tuning, and knowledge distillation—enables efficient and GDPR-compliant model updates. Using Florence-2 and the Hugging Face PEFT library, we demonstrate the framework on contamination object detection in organic waste streams across five distributed sites. Results show that the federated fusion model outperforms both centralized and local baselines in terms of IoU and detection accuracy, highlighting the effectiveness of LoRA-based adaptation for real-world, decentralized settings..

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Collaborative Trustworthy Foundation Model Framework: An Environmental Sustainability Use-Case to Detect Contamination Objects in Organic Waste Streams

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Alexander Valentinitsch,
  • Batuhan Bencik,
  • Mathias Brucker,
  • Gregor Lammer,
  • Cornelia Adami,
  • Mohit Kumar,
  • Lukas Fischer,
  • Florian Kromp

摘要

This work introduces a privacy-preserving framework that integrates foundation models with federated learning through a synergistic application using Low-Rank Adaptation (LoRA). Our three-stage pipeline—centralized pretraining, federated fine-tuning, and knowledge distillation—enables efficient and GDPR-compliant model updates. Using Florence-2 and the Hugging Face PEFT library, we demonstrate the framework on contamination object detection in organic waste streams across five distributed sites. Results show that the federated fusion model outperforms both centralized and local baselines in terms of IoU and detection accuracy, highlighting the effectiveness of LoRA-based adaptation for real-world, decentralized settings..