Robust Federated Learning: The Case of Affine Distribution Shifts

Authors: Amirhossein Reisizadeh, Farzan Farnia, Ramtin Pedarsani, Ali Jadbabaie

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We perform several numerical experiments to empirically support FLRA. We show that an affine distribution shift indeed suffices to significantly decrease the performance of the learnt classifier in a new test user, and our proposed algorithm achieves a significant gain in comparison to standard federated learning and adversarial training methods.
Researcher Affiliation Academia Amirhossein Reisizadeh ECE Department UC Santa Barbara reisizadeh@ucsb.edu Farzan Farnia MIT farnia@mit.edu Ramtin Pedarsani ECE Department UC Santa Barbara ramtin@ece.ucsb.edu Ali Jadbabaie MIT jadbabai@mit.edu
Pseudocode Yes Algorithm 1 Fed Robust
Open Source Code No The paper does not provide an unambiguous statement about releasing source code for their methodology or a link to a repository.
Open Datasets Yes We considered the standard MNIST [38] and CIFAR-10 [39] datasets and used three standard neural network architectures in the literature: Alex Net [40], Inception-Net [41], and a mini-Res Net [42].
Dataset Splits No The paper mentions 'm = 5000 training samples' and '5000 test samples', but does not specify a separate validation set split or its size.
Hardware Specification No The paper mentions that it was implemented in 'Tensorflow platform', but it does not specify any hardware details like CPU, GPU models, or memory.
Software Dependencies No The paper mentions 'Tensorflow platform [37]', but it does not provide a specific version number for Tensorflow or any other software dependencies.
Experiment Setup Yes In the experiments, we simulated a federated learning scenario with n = 10 nodes where each node observes m = 5000 training samples. We manipulated the training samples at each node via an affine distribution shift randomly generated according to a Gaussian distribution. We considered three standard neural network architectures in the literature: Alex Net [40], Inception-Net [41], and a mini-Res Net [42]. We tested the performance of the neural net classifiers trained by Fed Robust, Fed Avg, distributed FGM, and distributed PGD under different levels of affine distribution shifts. Figure 1 shows the accuracy performance over CIFAR-10 with Alex Net, Inception-Net, and Res Net architectures. We performed additional numerical experiments to analyze the effect of network size n and minimization iteration count on the robustness performance. Figure 3 shows the results of our experiments for a larger network size of n = 100 Alex Net neural network classifiers, each trained using m = 500 MNIST training data points. To examine the effect of parameter , i.e., minimization step count per training iteration, on our experimental results, we performed the CIFAR-10 experiment with the Alex Net architecture for = 5 as demonstrated in Figure 4.