Differentially Private and Communication Efficient Collaborative Learning

Authors: Jiahao Ding, Guannan Liang, Jinbo Bi, Miao Pan7219-7227

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

Reproducibility Variable Result LLM Response
Research Type Experimental The proposed methods are evaluated in extensive experiments on real-world datasets and the empirical results validate our theoretical findings.
Researcher Affiliation Academia 1University of Houston 2University of Connecticut {jding7, mpan2}@uh.edu, {guannan.liang, jinbo.bi}@uconn.edu
Pseudocode Yes Algorithm 1 Q-DPSGD-1 run by agent i, Algorithm 2 Q-DPSGD-2 run by agent i
Open Source Code No The paper does not provide any explicit statement about releasing code or a link to a code repository.
Open Datasets Yes We conduct the experiments over two benchmark datasets: MNIST and CIFAR-10.
Dataset Splits No The paper mentions training data ("randomly sample 10,000 records for training") and implicitly test data (through performance comparison figures), but does not explicitly describe a validation set or a three-way split for train/validation/test.
Hardware Specification No The paper does not provide specific details about the hardware used for running experiments, such as GPU or CPU models.
Software Dependencies No The paper mentions using Python for implementation (implied by typical ML papers) but does not provide specific version numbers for Python, libraries, or other software dependencies.
Experiment Setup Yes In the experiments, we set the step sizes (α, ε) = (0.3/T 1/6, 11/T 1/2) for Q-DPSGD-1 and Q-DPSGD-2, and α = 0.2 for DSGD and SDM. Moreover, we also set θ = 0.6 as stated in (Zhang et al. 2020) for SDM. To control the sensitivity of the gradient, we adopt gradient clipping threshold technique, ℓ(xi,t; θ) = ℓ(xi,t; θ)/ max (1, ℓ(xi,t; θ) /K). Here, we set K = 0.5 for Q-DPSGD-1 and Q-DPSGD-2 and SDM. In each simulation, we randomly sample 10,000 records for training and divide them into n parties, and thus each party consists of 10000/n data samples (i.e., m = 10000/n). In all experiments, we set δ = 10 5. We also set the processing speed of each machine follows a uniform distribution given as V Uniform(10, 90), and then choose the deadline Td = B/E[V ], where B is the expected batch size used in each machine. We consider a low precision quantizer in (5) with various quantization levels s, and we denote Tc as the communication time of a p-dimension vector without quantization (16 bits).