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; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1]
FedDr+: Stabilizing Dot-regression with Global Feature Distillation for Federated Learning
Authors: Seongyoon Kim, Minchan Jeong, Sungnyun Kim, Sungwoo Cho, Sumyeong Ahn, Se-Young Yun
TMLR 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our empirical results demonstrate that Fed Dr+ not only outperforms methods with a frozen classifier but also surpasses other state-of-the-art approaches, ensuring robust performance across diverse data distributions. We demonstrate the superiority of our method across various datasets and non-iid settings. Table 3 summarizes the accuracy comparison between various GFL methods proposed in recent literature and Fed Avg under various conditions. |
| Researcher Affiliation | Academia | Seongyoon Kim EMAIL Dept. ISys E, KAIST Minchan Jeong EMAIL KAIST AI Sungnyun Kim EMAIL KAIST AI Sungwoo Cho EMAIL KAIST AI Sumyeong Ahn EMAIL KENTECH, Department of Energy Engineering / Energy AI Se-Young Yun EMAIL KAIST AI |
| Pseudocode | Yes | The overall pseudocode of Fed Dr+ can be found in Appendix A. A.2 Pseudo Code of Fed Dr+ and Fed Dr+ FT We now present the pseudocode for Fed Dr+ and Fed Dr+ FT, outlining their key operations for global and personalized federated learning. The algorithm consists of two main stages: Algorithm 1 Fed Dr+, Fed Dr+ FT |
| Open Source Code | Yes | The code is available at: https://github.com/curisam/Fed Dr_plus. |
| Open Datasets | Yes | To simulate a realistic FL scenario involving 100 clients, we conduct extensive studies on three widely used datasets: CIFAR-10 (Krizhevsky et al., 2009), CIFAR-100 (Krizhevsky et al., 2009) and Image Net-100 (Deng et al., 2009). |
| Dataset Splits | Yes | The training data is distributed among the 100 clients using sharding and the LDA (Latent Dirichlet Allocation) partition strategies. Following the convention, sharding distributes the data into non-overlapping shards of equal size, each shard encompassing |Dtrain| / 100s and |Dtest| / 100s samples per class, where s denotes the number of shards per client. ... For Di test, we randomly sample from Dtest to match the class frequency of Di train and distribute it to each client i. |
| Hardware Specification | Yes | Our implementations are conducted using the Py Torch framework. Specifically, the experiments presented in Table 3 and Table 4 are executed on a single NVIDIA RTX 3090 GPU... The other parts of our study are carried out on a single NVIDIA A5000 GPU... |
| Software Dependencies | No | Our implementations are conducted using the Py Torch framework. Specifically, the experiments presented in Table 3 and Table 4 are executed on a single NVIDIA RTX 3090 GPU, based on the code structure from the following repository: https://github.com/Lee-Gihun/Fed NTD. The other parts of our study are carried out on a single NVIDIA A5000 GPU, utilizing the code framework from https://github.com/jhoon-oh/ Fed BABU. |
| Experiment Setup | Yes | In each round of communication, a random 10% of clients are selected to participate in the training process. The total number of communication rounds is set to 320. The initial learning rate and the number of local epochs for CIFAR-10, CIFAR-100, and Image Net-100 are determined through grid searches, with the detailed process and results provided in Appendix D. The learning rate η is decayed by a factor of 0.1 at the 160th and 240th communication rounds. Table 7: Summary of Dataset, Model, FL System, and Optimizer Specifications |