Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Byzantine-Robust Learning on Heterogeneous Datasets via Bucketing
Authors: Sai Praneeth Karimireddy, Lie He, Martin Jaggi
ICLR 2022 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We also theoretically and experimentally validate our approach, showing that combining bucketing with existing robust algorithms is effective against challenging attacks. |
| Researcher Affiliation | Academia | Sai Praneeth Karimireddy EMAIL Lie He EMAIL Martin Jaggi EMAIL |
| Pseudocode | Yes | Algorithm 1 Robust Aggregation (ARAGG) using bucketing; Algorithm 2 Robust Optimization using any Agnostic Robust Aggregator |
| Open Source Code | Yes | Implementations are based on Py Torch (Paszke et al., 2019) and will be made publicly available.2 The code is available at this url. https://github.com/epfl-lts2/byzantinerobust-noniid |
| Open Datasets | Yes | training an MLP on a heterogeneous version of the MNIST dataset (Le Cun et al., 1998). |
| Dataset Splits | No | The paper does not explicitly provide details about a validation set or specific training, validation, and test splits (e.g., 80/10/10). |
| Hardware Specification | No | The paper mentions training on "a cluster with 53 workers" multiple times, but does not specify any particular GPU models, CPU models, or other hardware specifications. |
| Software Dependencies | No | The paper states "Implementations are based on Py Torch (Paszke et al., 2019)" but does not provide a specific version number for PyTorch or any other software dependencies. |
| Experiment Setup | Yes | The results in Table 1, 3, 2, 4 are computed for 20 workers and 500 epochs with a learning rate of 0.01. The results in Figure 1 are computed for 25 workers, 600 epochs and with a learning rate of 0.05. We choose a small batch-size of 10 to highlight the challenges of heterogeneity. For CCLIP, we use a momentum coefficient of β = 0.9. For other aggregators, we choose β = 0.5. |