Federated Model Distillation with Noise-Free Differential Privacy

Authors: Lichao Sun, Lingjuan Lyu

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our extensive experimental results on various datasets validate that FEDMD-NFDP can deliver not only comparable utility and communication efficiency but also provide a noise-free differential privacy guarantee.
Researcher Affiliation Collaboration Lichao Sun1 , Lingjuan Lyu2 1 Lehigh University 2 Ant Group
Pseudocode Yes Algorithm 1 FEDMD-NFDP. Initialization phase does not involve collaboration. Di and wi are local dataset and model parameters from i-th party. Y i p[t] is the prediction from i-th party on the chosen public subset Xp Dp in round t.
Open Source Code No The paper states that experiments are implemented using PyTorch but does not provide any explicit statement about open-sourcing its own code or a link to a repository.
Open Datasets Yes In the experiment, we evaluate on paired datasets, i.e., MNIST/FEDMNIST and CIFAR-10/CIFAR-100. For MNIST/FEMNIST, the public data is the MNIST, and the private data is a subset of the Federated Extended MNIST (FEMNIST) [Caldas et al., 2018]... For CIFAR-10/CIFAR-100, the public dataset is the CIFAR-10, and the private dataset is a subset of the CIFAR-100 [Sun et al., 2021].
Dataset Splits No The paper mentions training on 'sampled subset from private local data' and 'unlabeled public set' but does not provide specific details on train/validation/test splits, percentages, or sample counts, nor does it explicitly refer to a 'validation' set for model selection.
Hardware Specification Yes A single GPU NVIDIA Tesla V100 is used in the experiments.
Software Dependencies No The paper states 'All experiments are implemented by using Pytorch' but does not specify a version number for PyTorch or any other software dependencies.
Experiment Setup Yes In each communication round, we use a subset of size 5000 that is randomly selected from the entire public dataset... We empirically choose R = 20, T1 = 20, T2 = 2, T3 = 1 via grid search. We initialize all parties with the same pre-trained model on some labelled data in the same domain. and Each party s local model is two or three-layer deep neural networks for both MNIST/FEMNIST and CIFAR-10/CIFAR-100.