Game of Gradients: Mitigating Irrelevant Clients in Federated Learning

Authors: Lokesh Nagalapatti, Ramasuri Narayanam9046-9054

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

Reproducibility Variable Result LLM Response
Research Type Experimental We finally conduct a thorough empirical analysis on image classification and speech recognition tasks to show the superior performance of S-Fed Avg than the baselines in the context of supervised federated learning settings. Extensive experimentation that shows the efficacy of our solution approach.
Researcher Affiliation Collaboration Lokesh Nagalapatti*,1 Ramasuri Narayanam2 1 IIT Bombay, Mumbai India 2 IBM Research India nlokeshiisc@gmail.com, ramasurn@in.ibm.com
Pseudocode Yes Algorithm 1: Shapley values based Federated Averaging (S-Fed Avg) Algorithm 2: Compute Shapley values of (δt+1, v)
Open Source Code Yes 2Supplementary material: https://github.com/nlokeshiisc/SFedAvg-AAAI21
Open Datasets Yes The MNIST dataset is composed of handwritten digits. 1MNIST: http://yann.lecun.com/exdb/mnist/
Dataset Splits Yes In both the cases, server contains a dataset of about 5000 even digits which is further partitioned into DV , DT est such that |DV | = 1000 and |DT est| = 4000 respectively.
Hardware Specification No The paper does not provide specific details regarding the hardware used for running the experiments (e.g., GPU models, CPU types, or cloud computing specifications).
Software Dependencies No The paper mentions using stochastic gradient descent and neural networks, but it does not specify any software libraries or frameworks (e.g., TensorFlow, PyTorch) along with their version numbers that are necessary for replication.
Experiment Setup Yes Hyper-Parameter Values: The values of hyperparameters that we use in the experiments are: K = 10, m = 5, B = 32, α = 0.75, γ = 0.95, λ = 2%, β = 0.25, ζ = 5, R = 10, ηi = 0.01. In all our experiments, we set ηk = 0.01 and E = 5.