FOOGD: Federated Collaboration for Both Out-of-distribution Generalization and Detection
Authors: Xinting Liao, Weiming Liu, Pengyang Zhou, Fengyuan Yu, Jiahe Xu, Jun Wang, Wenjie Wang, Chaochao Chen, Xiaolin Zheng
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In empirical validations, FOOGD significantly enjoys three main advantages: (1) reliably estimating nonnormalized decentralized distributions, (2) detecting semantic shift data via score values, and (3) generalizing to covariate-shift data by regularizing feature extractor. The prejoct is open in https://github.com/Xenia LLL/FOOGD-main.git. and 5 Experiments |
| Researcher Affiliation | Collaboration | Xinting Liao1, Weiming Liu1, Pengyang Zhou1, Fengyuan Yu1, Jiahe Xu1, Jun Wang2, Wenjie Wang3, Chaochao Chen1, Xiaolin Zheng1 1 Zhejiang University, 2 OPPO Research Institute, 3 National University of Singapore |
| Pseudocode | Yes | A Algorithms The overall algorithm of FOOGD is in Algo. 1. In line 1:10, the server collaborates with clients to optimize the feature extractor model for representation and the score model for density estimation. The clients execute training local models separately in line 11:19. In each client, SM3D estimates data density based on the latent representation of feature extractor, and SAG computes the kernelized stein discrepancy based on score model to regularize the optimization of feature extractor. We update score model with SM3D in Algo. 2, and the training procedure of feature extractor is detailed in Algo. 3. |
| Open Source Code | Yes | The prejoct is open in https://github.com/Xenia LLL/FOOGD-main.git. |
| Open Datasets | Yes | Datasets. Following SCONE [3], we choose clear Cifar10, Cifar100 [33], and Tiny Image Net [36] as the IN data, and select the corresponding corrupted versions [19], i.e., Cifar10-C, Cifar100-C and Tiny Image Net-C as IN-C data. We evaluate detection with five OUT image datasets: SVHN [59], Texture [11], i SUN [78], LSUN-C and LSUN-R [81]. To simulate the non-IID scenarios, we sample data by label in a Dirichlet distribution parameterized by non-IID degree [23], i.e., α, for K clients. The smaller α simulates the more heterogeneous client data distribution in federated settings. To evaluate FOOGD on unseen client generalization data, we also use PACS [38] dataset for leave-one-out domain generalization. |
| Dataset Splits | No | The paper mentions 'validation' in section 5.1, but does not provide specific details on the dataset split for training, validation, and testing, such as percentages or sample counts. It only states 'The original train and test datasets are split to all clients'. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU models, CPU types, or memory) used for running the experiments. The justification in the NeurIPS checklist states 'My research group supports me in computer resources' but offers no specifics. |
| Software Dependencies | No | The paper mentions using 'SGD optimizer' and specific neural network architectures ('Wide Res Net', 'Res Net18'), but it does not specify versions for any software libraries or frameworks (e.g., PyTorch, TensorFlow, or specific Python packages) that are crucial for reproducibility. |
| Experiment Setup | Yes | Implementation Details. We choose Wide Res Net [85] as our main task model for Cifar datasets, and Res Net18 [18] for Tiny Image Net and PACS, and optimize each model 5 local epochs per communication round until converging with SGD optimizer. We conduct all methods at their best and report the average results of three repetitions with different random seeds. We consider client number K = 10, participating ratio of 1.0 for performance comparison, and the hyperparameters λm = 0.5, λa = 0.05. |