FedCache

This repository is the official Pytorch implementation DEMO of FedCache: A Knowledge Cache-driven Federated Learning Architecture for Personalized Edge Intelligence. IEEE Transactions on Mobile Computing (TMC). 2024

News

[Apr. 2024] FedCache's remote information retrieval has been effectively advanced and implemented by PTaaS. Privacy-Preserving Training-as-a-Service for On-Device Intelligence: Concept, Architectural Scheme, and Open Problems (arxiv.org).

[Mar. 2024] FedCache is featured by Tencent. 机器人再度大幅进化！阿西莫夫三法则还有效吗？(Robots are Evolving Dramatically Again! Is Asimov's "Three Laws of Robotics" Still Valid?).

[Mar. 2024] I was invited to give a talk for Network System and Machine Learning Group, School of Computer Science, Peking University. 面向个性化边缘智能的缓存驱动联邦学习: 研究进展与开放性问题 (Cache-driven Federated Learning for Personalized Edge Intelligence: Research Progress and Open Problems).

[Mar. 2024] FedCache is featured by NGUI. 缓存驱动联邦学习架构赋能个性化边缘智能 (Cache-Driven Federated Learning Architecture Energizes Personalized Edge Intelligence).

[Mar. 2024] FedCache is included by the first survey investigating the application of knowledge distillation in federated edge learning. Knowledge Distillation in Federated Edge Learning: A Survey (arxiv.org).

[Feb. 2024] FedCache is featured on Phoenix Tech. 缓存驱动联邦学习架构来了！专为个性化边缘智能打造 (The Cache-Driven Federated Learning Architecture is Coming! Built for Personalized Edge Intelligence).

[Feb. 2024] FedCache is accepted by IEEE Transactions on Mobile Computing (TMC). FedCache: A Knowledge Cache-Driven Federated Learning Architecture for Personalized Edge Intelligence | IEEE Journals & Magazine | IEEE Xplore

[Jan. 2024] One follow-up paper examines the impact of logits poisoning attack on FedCache. Logits Poisoning Attack in Federated Distillation (arxiv.org).

[Dec. 2023] We discover a Chinese blog that interprets FedCache on CSDN. 缓存驱动的联邦学习架构FedCache (FedCache: Cache-Driven Federated Learning Architecture).

[Dec. 2023] One follow-up paper confirms the further potential of FedCache for enhanced communication efficiency by accumulating local updates. Improving Communication Efficiency of Federated Distillation via Accumulating Local Updates (arxiv.org).

[Aug. 2023] FedCache is featured by Netease. AI在量子计算中的研究进展 (Research Progress of AI in Quantum Computing).

[Aug. 2023] FedCache is released on arxiv. FedCache: A Knowledge Cache-driven Federated Learning Architecture for Personalized Edge Intelligence (arxiv.org).

Highlight

FedCache is a device friendly, scalable and effective personalized federated learning architecture tailored for edge computing.
FedCache guarantees satisfactory performance while conforming to multiple personalized devices-side limitations.
FedCache improves communication efficiency by up to x200 over previous architectures and can accommodate heterogeneous devices and asynchronous interactions among devices and the server.

Family of FedCache

Foundation Works: FedICT, FedDKC, MTFL, DS-FL, FD
Derivative Works:
- Communication: ALU
- Poisoning Attack: FDLA
- Generalization: Coming Soon......
- Security: Coming Soon......
- Application: Coming Soon......
- Robustness: TBD
- Scaling: TBD
- Fairness: TBD
- Deployment: TBD

If you have any ideas or questions regarding to FedCache, please feel free to contact [email protected].

Requirements

Python: 3.10
Pytorch: 1.13.1
torchvision: 0.14.1
hnswlib
Other dependencies

Run this DEMO

python main_fedcache.py

Evaluation

Model Homogeneous Setting

MNIST Dataset

Method	MAUA (%)	Communication Cost (G)	Speed-up Ratio
pFedMe	94.89	13.25	×1.0
MTFL	95.59	7.77	×1.7
FedDKC	89.62	9.13	×1.5
FedICT	84.62	-	-
FD	84.19	-	-
FedCache	87.77	0.99	×13.4

FashionMNIST Dataset

Method	MAUA (%)	Communication Cost (G)	Speed-up Ratio
pFedMe	81.57	20.71	×1.0
MTFL	83.92	12.33	×1.7
FedDKC	78.24	8.43	×2.5
FedICT	76.90	13.34	×1.6
FD	76.32	-	-
FedCache	77.71	0.08	×258.9

CIFAR-10 Dataset

Method	MAUA (%)	Communication Cost (G)	Speed-up Ratio
pFedMe	37.49	-	-
MTFL	43.43	52.99	×1.0
FedDKC	45.87	11.46	×4.6
FedICT	43.61	10.69	×5.0
FD	42.77	-	-
FedCache	44.42	0.19	×278.9

CINIC-10 Dataset

Method	MAUA (%)	Communication Cost (G)	Speed-up Ratio
pFedMe	31.65	-	-
MTFL	34.09	-	-
FedDKC	43.95	4.12	×1.3
FedICT	42.79	5.50	×1.0
FD	39.36	-	-
FedCache	40.45	0.07	×78.6

Model Heterogeneous Setting

MNIST Dataset

Method	MAUA (%)	Communication Cost (G)	Speed-up Ratio
FedDKC	85.38	10.53	×1.0
FedICT	80.53	-	-
FD	79.90	-	-
FedCache	83.94	0.10	×105.3

FashionMNIST Dataset

Method	MAUA (%)	Communication Cost (G)	Speed-up Ratio
FedDKC	77.96	12.64	×1.0
FedICT	76.11	-	-
FD	75.57	-	-
FedCache	77.26	0.08	×158.0

CIFAR-10 Dataset

Method	MAUA (%)	Communication Cost (G)	Speed-up Ratio
FedDKC	44.53	4.58	×1.2
FedICT	43.96	5.35	×1.0
FD	40.40	-	-
FedCache	41.59	0.05	×107.0

CINIC-10 Dataset

Method	MAUA (%)	Communication Cost (G)	Speed-up Ratio
FedDKC	44.80	4.12	×1.3
FedICT	43.40	5.50	×1.0
FD	40.76	-	-
FedCache	41.71	0.07	×78.6

Cite this work

@ARTICLE{wu2024fedcache,
  author={Wu, Zhiyuan and Sun, Sheng and Wang, Yuwei and Liu, Min and Xu, Ke and Wang, Wen and Jiang, Xuefeng and Gao, Bo and Lu, Jinda},
  journal={IEEE Transactions on Mobile Computing}, 
  title={FedCache: A Knowledge Cache-Driven Federated Learning Architecture for Personalized Edge Intelligence}, 
  year={2024},
  volume={},
  number={},
  pages={1-15},
  keywords={Computer architecture;Training;Servers;Computational modeling;Data models;Adaptation models;Performance evaluation;Communication efficiency;distributed architecture;edge computing;knowledge distillation;personalized federated learning},
  doi={10.1109/TMC.2024.3361876}
  }