Towards the Robustness of Differentially Private Federated Learning

Authors: Tao Qi, Huili Wang, Yongfeng Huang

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on three benchmark datasets demonstrate that baseline methods cannot ensure task accuracy, data privacy, and robustness simultaneously, while Robust-DPFL can effectively enhance the privacy protection and robustness of federated learning meanwhile maintain the task performance.
Researcher Affiliation Academia 1Department of Electronic Engineering, Tsinghua University, Beijing 100084, China 2Zhongguancun Laboratory, Beijing 100094, China
Pseudocode No The paper describes its methods and algorithms using prose and mathematical equations but does not include any formal pseudocode blocks or algorithm listings.
Open Source Code Yes The code is available in https://github.com/taoqi98/Robust-DPFL.
Open Datasets Yes We conduct experiments on three benchmark datasets for federated learning, including MNIST (Deng 2012), FEMNIST (Caldas et al. 2018), and CIFAR-10 (Krizhevsky, Hinton et al. 2009).
Dataset Splits No The paper describes how training data is partitioned among clients ('The training data is randomly partitioned into 100 clients based on a non-IID data partition strategy'), but it does not specify a distinct validation set split or its size for hyperparameter tuning or early stopping.
Hardware Specification No The paper does not specify the hardware used for experiments, such as CPU or GPU models, or memory specifications.
Software Dependencies No The paper mentions that models are implemented by ResNet-18, but does not list specific software versions (e.g., Python, PyTorch, TensorFlow) or library dependencies with version numbers.
Experiment Setup Yes We conduct experiments on three benchmark datasets for federated learning, including MNIST (...), FEMNIST (...), and CIFAR-10 (...). The training data is randomly partitioned into 100 clients based on a non-IID data partition strategy (...). The basic machine learning models trained on these three datasets are implemented by ResNet-18 (...). The proportion of malicious clients controlled by the adversary is set to 15%. ... The level of DP guarantee is set to (1.2, 5). ... For a malicious client u, Attack DPFL first learns the poisoned gradient Gu from its local poisoned data. Then Attack-DPFL simulates the distribution of the perturbed poisoned gradient and aligns their norms to learn an amplified unperturbed poisoned gradient Au: Au = AGu, A = ||Su||/||Gu|| Su = h(Gu; l) + N(0, σ2), (4) where A is the amplification coefficient. ... In each training round of Robust-DPFL, the server first models the detection score of each uploaded gradient {Z(Su)|u Ut}. Then the server clusters the detection scores into two groups based on the K-means algorithm to detect the suspicious gradients.