The Distributed Discrete Gaussian Mechanism for Federated Learning with Secure Aggregation

Authors: Peter Kairouz, Ziyu Liu, Thomas Steinke

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

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically evaluate our protocol and compare it to the commonly used centralized continuous Gaussian mechanism on two canonical tasks: distributed mean estimation (DME) and federated learning (FL). ... Figure 1 shows the mean MSE x bx 2 2/d with 95% confidence interval over 10 random dataset initializations. ... Figure 2 shows the test accuracies on EMNIST and SO-NWP respectively.
Researcher Affiliation Industry Peter Kairouz 1 Ziyu Liu 1 Thomas Steinke 1 ... 1Google Research. Correspondence to: Peter Kairouz <kairouz@google.com>, Ziyu Liu <klz@google.com>, Thomas Steinke <ddg@thomas-steinke.net>.
Pseudocode Yes Algorithm 1 Client Procedure Aclient
Open Source Code Yes We provide an open-source implementation of our methods in Tensor Flow Privacy (Andrew et al., 2019) and Tensor Flow Federated (Ingerman & Ostrowski, 2019).4 ... 4Code: https://github.com/google-research/federated/tree/master/distributed_dp.
Open Datasets Yes For FL, we show on Federated EMNIST (Caldas et al., 2018) and Stack Overflow (Authors, 2019) that our approach gives good performance under tight privacy budgets... Authors, T. T. F. Tensorflow federated stack overflow dataset. 2019. URL https://www.tensorflow.org/ federated/api_docs/python/tff/ simulation/datasets/stackoverflow/ load_data.
Dataset Splits No The paper mentions 'validation accuracies' in Figure 3, indicating a validation process was performed. However, it does not explicitly state the specific dataset split percentages (e.g., 'X% training, Y% validation, Z% test') or the methodology for creating the validation set itself from the overall dataset to enable reproduction of data partitioning. It primarily describes client sampling and training rounds.
Hardware Specification No The paper does not provide specific details about the hardware used for its experiments, such as GPU/CPU models, memory, or cloud computing instance types.
Software Dependencies No The paper states, 'We provide an open-source implementation of our methods in Tensor Flow Privacy (Andrew et al., 2019) and Tensor Flow Federated (Ingerman & Ostrowski, 2019).' However, it does not specify version numbers for these or any other software dependencies, which is required for reproducible description.
Experiment Setup Yes For both tasks, we train with federated averaging with server momentum of 0.9... In each round, we uniformly sample n = 100 clients... For EMNIST and SO-NWP respectively, we set the number of rounds T to 1500 and 1600, c to 0.03 and 0.3, client learning rate ηc to 0.032 and 0.5, and client batch size to 20 and 16. Server LR ηs is set to 1 for EMNIST and selected from a small grid {0.3, 1} for SO-NWP.