Sageflow: Robust Federated Learning against Both Stragglers and Adversaries

Authors: Jungwuk Park, Dong-Jun Han, Minseok Choi, Jaekyun Moon

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experimental results show that Sageflow outperforms various existing methods aiming to handle stragglers/adversaries. 4 Experiments In this section, we validate Sageflow on MNIST [10], FMNIST [23] and CIFAR10 [8].
Researcher Affiliation Academia Jungwuk Park KAIST savertm@kaist.ac.kr Dong-Jun Han KAIST djhan93@kaist.ac.kr Minseok Choi Jeju National University ejaqmf@jejunu.ac.kr Jaekyun Moon KAIST jmoon@kaist.edu
Pseudocode Yes Algorithm 1 Proposed Sageflow Algorithm
Open Source Code No The paper does not provide a direct link to open-source code or explicitly state that the code for the methodology is being released.
Open Datasets Yes We validate Sageflow on MNIST [10], FMNIST [23] and CIFAR10 [8]. The dataset is split into 60,000 train and 10,000 test samples for MNIST and FMNIST, and split into 50,000 train and 10,000 test samples for CIFAR10.
Dataset Splits No The paper specifies training and test splits, for example: 'The dataset is split into 60,000 train and 10,000 test samples for MNIST and FMNIST, and split into 50,000 train and 10,000 test samples for CIFAR10.' However, it does not explicitly mention a separate validation dataset split with specific numbers or percentages for the main experiments.
Hardware Specification No The paper mentions 'edge devices' and 'server' but does not specify any particular hardware components like GPU models, CPU types, or memory specifications used for running the experiments.
Software Dependencies No The paper mentions using a 'convolutional neural network (CNN)' and 'VGG-11' but does not specify any software dependencies (e.g., Python, TensorFlow, PyTorch) with their version numbers.
Experiment Setup Yes The number of local epochs at each device is set to 5. The local batch size is set to 10 for all experiments except for the backdoor attack. At each global round, we randomly selected a fraction C of devices in the system to participate. For the proposed Sageflow method, we sample 2% of the entire training data uniformly at random to be the public data... γ is decayed while the learning rate is decayed in other schemes. For model poisoning, each adversarial device sends 0.1w to the server... For data poisoning attack, we conduct label-flipping... For both attacks, we set C to 0.2 and the portion of adversarial devices is assumed to be r = 0.2 at each global round.