Core Concepts
SFLにおけるカットレイヤー選択は、再構築攻撃のリスクを最小限に抑えながら、必要なエネルギーバジェット内でエネルギー消費を維持するために重要です。
Abstract
この記事では、Split Federated Learning(SFL)のプロセス全体とカットレイヤーの選択がSFLに与える影響について包括的な概要が提供されています。主な焦点は、エネルギー消費とプライバシーへの影響です。以下は内容の詳細な概要です:
Abstract:
- SFL combines federated learning (FL) and split learning (SL).
- Cut layer selection in SFL impacts energy consumption and privacy.
I. INTRODUCTION:
- FL distributes training process across clients.
- SL breaks down DL model into sub-models.
- SFL allows parallel local training while engaging with servers.
II. WHY IS CUT LAYER SELECTION IMPORTANT?:
- Cut layer influences energy consumption and privacy.
- Selection strategy crucial for minimizing reconstruction attacks.
III. CASE STUDY: CUT LAYER SELECTION:
- Energy consumption model formulated based on client interactions.
- Privacy risks evaluated through SSIM between original and reconstructed images.
IV. OPEN CHALLENGES:
- Deep Reinforcement Learning for optimal cut layer selection.
- Privacy protection against label inference attacks.
- Lightweight design through quantization approach.
V. CONCLUSIONS:
- Importance of cut layer selection for balancing privacy and energy consumption in SFL.