Toward Understanding the Influence of Individual Clients in Federated Learning
Authors: Yihao Xue, Chaoyue Niu, Zhenzhe Zheng, Shaojie Tang, Chengfei Lyu, Fan Wu, Guihai Chen10560-10567
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical results on a synthetic dataset and the FEMNIST dataset demonstrate that our estimation method can approximate Fed-Influence with small bias. Further, we show an application of Fed-Influence in model debugging. |
| Researcher Affiliation | Collaboration | 1 Shanghai Jiao Tong University, 2 The University of Texas at Dallas, 3 Alibaba Group |
| Pseudocode | No | The paper mentions an algorithm in the supplement: |
| Open Source Code | No | The paper states: |
| Open Datasets | Yes | Empirical results on a synthetic dataset and the FEMNIST dataset (Caldas et al. 2018) demonstrate the effectiveness of our method. We used the softmax function at the output layer and adopted the cross entropy as the loss function. In setting 1, the loss function is convex but not strongly convex, and therefore it is in Case 2 (γ = 1). In setting 2, although the toy model has no activation function, which makes it equivalent to a single-layer perceptron with convex loss function, results show that it is still in Case 3 (γ > 1) because the learning rate is too large. And in setting 3, the loss function is non-convex and is therefore in Case 3, too. Table 2: Detailed configuration of the three different settings. The two datasets are described in Caldas et al. (2018). |
| Dataset Splits | No | The paper refers to a |
| Hardware Specification | Yes | Experiments are conducted on 64bit Ubuntu 18.04 LTS with four Intel i9-9900K CPUs and two NVIDIA RTX-2080TI GPUs, 200GB storage. |
| Software Dependencies | No | The paper mentions using |
| Experiment Setup | Yes | We take Leaf (Caldas et al. 2018), a benchmarking framework for federated learning based on tensorflow. We evaluated our method on three settings, as shown in Table 2. We used the softmax function at the output layer and adopted the cross entropy as the loss function. Model Dataset Distribution η |C| |Ct| m T Ns Setting 1 Log Reg Synthetic Non-IID, Unbalanced 0.003 1000 10 5 1000 50 Setting 2 CNN 1 FEMNIST IID, Balanced 0.03 50 5 2 500 50 Setting 3 CNN 2 FEMNIST Non-IID, Unbalanced 0.02 100 10 2 2000 50 |