Library

FOOGD: Federated Collaboration for Both Out-of-distribution Generalization and Detection

Federated learning (FL) is a promising machine learning paradigm that collaborates with client models to capture global knowledge. However, deploying FL models in real-world scenarios remains unreliable due to the coexistence of in-distribution data and unexpected out-of-distribution (OOD) data, such as covariate-shift and semantic-shift data. Current FL researches typically address either covariate-shift data through OOD generalization or semantic-shift data via OOD detection, overlooking the simultaneous occurrence of various OOD shifts. In this work, we propose FOOGD, a method that estimates the probability density of each client and obtains reliable global distribution as guidance for the subsequent FL process. Firstly, SM3D in FOOGD estimates score model for arbitrary distributions without prior constraints, and detects semantic-shift data powerfully. Then SAG in FOOGD provides invariant yet diverse knowledge for both local covariate-shift generalization and client performance generalization. In empirical validations, FOOGD significantly enjoys three main advantages: (1) reliably estimating non-normalized 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/XeniaLLL/FOOGD-main.git.

cs.LGAnomaly Detection Techniques and ApplicationsComputer Science
View graphSource

Paper

References (90)

01CE-RCFR: Robust Counterfactual Regression for Consensus-Enabled Treatment Effect EstimationFan Wang, Chaochao Chen, Weiming Liu et al.2024 · Knowledge Discovery and Data Mining · 23 citations
02Counterfactual User Sequence Synthesis Augmented with Continuous Time Dynamic Preference Modeling for Sequential POI RecommendationLianyong Qi, Yuwen Liu, Weiming Liu et al.2024 · International Joint Conference on Artificial Intelligence · 12 citations
03Envisioning Outlier Exposure by Large Language Models for Out-of-Distribution DetectionChentao Cao, Zhun Zhong, Zhanke Zhou et al.2024 · International Conference on Machine Learning · 39 citations In Library
04Out-of-distribution Detection Learning with Unreliable Out-of-distribution SourcesHaotian Zheng, Qizhou Wang, Zhen Fang et al.2023 · Neural Information Processing Systems · 51 citations In Library
05Learning to Augment Distributions for Out-of-Distribution DetectionQizhou Wang, Zhen Fang, Yonggang Zhang et al.2023 · Neural Information Processing Systems · 56 citations In Library
06Contrastive Proxy Kernel Stein Path Alignment for Cross-Domain Cold-Start RecommendationWeiming Liu, Xiaolin Zheng, Jiajie Su et al.2023 · IEEE Transactions on Knowledge and Data Engineering · 25 citations
07Sampling Multimodal Distributions with the Vanilla Score: Benefits of Data-Based InitializationFrederic Koehler, T. Vuong2023 · International Conference on Learning Representations · 11 citations In Library
08Joint Local Relational Augmentation and Global Nash Equilibrium for Federated Learning with Non-IID DataXinting Liao, Chaochao Chen, Weiming Liu et al.2023 · ACM Multimedia · 11 citations In Library
09Variational Inference with Gaussian Score MatchingC. Modi, Charles C. Margossian, Yuling Yao et al.2023 · Neural Information Processing Systems · 23 citations In Library
10L-DAWA: Layer-wise Divergence Aware Weight Aggregation in Federated Self-Supervised Visual Representation LearningYasar Abbas Ur Rehman, Yan Gao, Pedro Gusmão et al.2023 · IEEE International Conference on Computer Vision · 47 citations In Library
11Feed Two Birds with One Scone: Exploiting Wild Data for Both Out-of-Distribution Generalization and DetectionHaoyue Bai, Gregory H. Canal, Xuefeng Du et al.2023 · International Conference on Machine Learning · 72 citations In Library
12DOS: Diverse Outlier Sampling for Out-of-Distribution DetectionWenyu Jiang, Hao Cheng, Mingcai Chen et al.2023 · International Conference on Learning Representations · 19 citations In Library

Scroll for more · 38 remaining

Similar papers

Federated Graph Out-of-Distribution Generalization via Representation Propagation and ScatteringYukai Zhu, He Sun, Mingjun Xiao2025 In Library
FedGOG: Federated Graph Out-of-Distribution Generalization with Diffusion Data Exploration and Latent Embedding DecorrelationPengyang Zhou, Chaochao Chen, Weiming Liu et al.2025 · Proceedings of the AAAI Conference on Artificial Intelligence · 9 citations In Library
Unified OOD Detection and Generalization for Federated Learning with Data HeterogeneityLiang Lu, Xiaodong Fu, Lianyin Jia2025 In Library
Adaptive Dual-Weighting Framework for Federated Learning via Out-of-Distribution DetectionZhiwei Ling, Hailiang Zhao, Chaoyu Zhang et al.2026 · arXiv.org · 0 citations In Library
A study about the performance of federated learning under out-of-distribution data groupsMinghao Wang2025 · Springer Link (Chiba Institute of Technology) In Library
A study about the performance of federated learning under out-of-distribution data groupsMinghao Wang2025 · ITM Web of Conferences In Library
OLG-FL: A Federated Learning Framework for Optimizing Local Training and Global AggregationHaotong Qiu, Zhenyan Liu, Xiaolin Zhao et al.2025 In Library
Federated Domain Generalization with Data-free On-server Matching GradientTrong-Binh Nguyen, D. Nguyen, Jinsun Park et al.2025 · International Conference on Learning Representations · 13 citations In Library
FL Games: A Federated Learning Framework for Distribution ShiftsSharut Gupta, Kartik Ahuja, Mohammad Havaei et al.2022 · arXiv (Cornell University) In Library
<b>Fed-OGD:</b> Mitigating Straggler Effects in Federated Learning via Orthogonal Gradient DescentWei Li, Zicheng Shen, Xiulong Liu et al.2025 · IEEE Transactions on Computers · 7 citations In Library

Peer review

Accept (poster)NeurIPS 2024 · via OpenReview
Reviewer dEK55/10 · confidence 4/52024-06-15

Summary

This paper introduces a FL framework FOOGD aimed at addressing the simultaneous challenges of the out-of-distribution (OOD) generalization and OOD detection in decentralized environments. FOOGD estimates the data distribution of different clients through SM3D and introduces SAG to ensure the consistency of features.

Strengths

FOOGD is designed to handle both OOD generalization and OOD detection tasks and integrates mechanisms to manage covariate shifts and semantic shifts within the FL setup, which is helpful for deployment environments. The use of SM3D allows for accurate estimation of the probability density of client data distributions and is used to identify out-of-distribution samples. In addition, SAG helps maintain feature space invariance between in-distribution and augmented data samples. Finally, this paper demonstrates the effectiveness of FOOGD through experiments.

Weaknesses

There is no training-log provided in supp materials. There are also questions regarding the foundation methods as listed below.

Questions

(1) The dual approach of handling both OOD generalization and detection introduces significant computational complexity and overhead. However, FOOGD is not clearly stated, and there are no experiments to quantitatively analyze it. The iterative process of updating and aggregating both main task models and score models can still be costly in terms of bandwidth and synchronization requirements among clients. (2) The need to transmit detailed model parameters and score function values between clients and the server can lead to high communication costs. In addition, the score function values may cause privacy leakage issues, thereby attacking the score model. (3) While the FOOGD framework aims to generalize across different client distributions, the inherent diversity and possible sparsity of client data can cause the model to overly adapt to specific local patterns, reducing its effectiveness in capturing broader, global trends. (4) The effectiveness of FOOGD heavily relies on the accuracy of the local density estimation provided by the SM3D component. However, in scenarios where data distributions are highly complex or sparse, it may lead to less reliable detection of OOD samples and potential false positives or negatives.

Rating

5

Confidence

4

Soundness

3

Presentation

2

Contribution

2

Limitations

The authors indicated the paper's focus (on privacy issues) as an limitation but not elaborated in paper itself.

Reviewer cFix5/10 · confidence 4/52024-07-08

Summary

This paper introduces FOOGD, a federated collaboration framework designed to achieve both out-of-distribution (OOD) generalization and detection. FOOGD estimates the probability density of each client to obtain a reliable global distribution and incorporates the SM3D model and SAG module to enhance the detection capability of the score model and the generalization ability of the feature extractor.

Strengths

1. FOOGD is the first framework in federated learning to simultaneously address OOD generalization and detection, demonstrating significant innovation. 2. The paper is written clearly and coherently, and it is generally well-organized. 3. The authors have made an effort to highlight their contributions, providing a reasonably good overview of the key points and innovations presented in the paper.

Weaknesses

1. The experimental section discusses the results and performance comparisons. Although this section provides performance metrics for multiple datasets, the description of the chosen statistical analysis methods (such as whether hypothesis testing or confidence interval calculation was performed) is insufficient, affecting the reliability of the result interpretation. 2. Evaluation metrics like ACC-IN and ACC-IN-C are not explained. 3. The related work section only introduces some comparison methods and lacks more related work, such as FedAvg. Besides, some references are incomplete, such as ref. 46. 4. No open-source code has been provided.

Questions

1. Could the experimental results section provide a more detailed analysis of OOD to better explain the superiority of the FOOGD method in this aspect?

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

See the above weakness.

Authorsrebuttal2024-08-10

Response to Reviwer cFix

**We are not planning to address the problems you raised in future work, instead, we reverse these problems in our current work. The future version indicates that the camera-ready version we revised based on the current version, because it is impossible to resubmit on the current version.** We elaborate our revised version in the following. **The evidential results for weakness 1 is explained in our A1, where we provide the reason for choosing accuracy for OOD generalization, and FPR95 and AUROC for OOD detection, which statistically follow the work of FedTHE, FOSTER, and SCONE.** This is introduced in lines 272-282 of main paper, and lines 654-663 in appendix. We omit the statistics in our rebuttal due to space limits. And we provide the revised two main tables as below. **Table 1: Main results of federated OOD detection and generalization on Cifar10. We report the ACC of brightness as IN-C ACC, the FPR95 and AUROC of LSUN-C as OUT performance.** | Non-IID\Method | $\alpha=0.1$ | | | | $\alpha=0.5$ | | | | $\alpha=5$ | | | | | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | | | ACC-IN $\uparrow$ | ACC-IN-C $\uparrow$ | FPR95 $\downarrow$ | AUROC $\uparrow$ | ACC-IN $\uparrow$ | ACC-IN-C $\uparrow$ | FPR95 $\downarrow$ | AUROC $\uparrow$ | ACC-IN $\uparrow$ | ACC-IN-C $\uparrow$ | FPR95 $\downarrow$ | AUROC $\uparrow$ | | FedAvg | 68.03 $\pm$ 1.17 | 65.44 $\pm$ 1.18 | 83.41 $\pm$ 1.57 | 58.05 $\pm$ 0.89 | 86.59 $\pm$ 1.13 | 83.72 $\pm$ 1.74 | 43.70 $\pm$ 0.83 | 84.18 $\pm$ 0.23 | 86.50 $\pm$ 0.33 | 85.08 $\pm$ 0.49 | 38.24 $\pm$ 0.55 | 85.37 $\pm$ 0.29 | | FedLN | 75.24 $\pm$ 0.44 | 71.77 $\pm$ 0.67 | 56.14 $\pm$ 0.91 | 84.14 $\pm$ 0.37 | 86.10 $\pm$ 0.89 | 84.20 $\pm$ 1.82 | 39.26 $\pm$ 1.14 | 89.64 $\pm$ 0.52 | 87.20 $\pm$ 1.26 | 85.08 $\pm$ 1.43 | 33.33 $\pm$ 2.38 | 90.87 $\pm$ 0.58 | | FedATOL | 55.93 $\pm$ 1.87 | 54.44 $\pm$ 1.72 | 49.50 $\pm$ 1.59 | 86.22 $\pm$ 2.74 | 87.55 $\pm$ 0.91 | 85.64 $\pm$ 0.54 | 27.87 $\pm$ 1.32 | 93.48 $\pm$ 0.69 | 89.27 $\pm$ 0.68 | 88.28 $\pm$ 1.32 | 19.66 $\pm$ 2.62 | 95.25 $\pm$ 0.78 | | FedT3A | 68.03 $\pm$ 1.17 | 61.52 $\pm$ 1.39 | 83.41 $\pm$ 1.57 | 58.05 $\pm$ 0.89 | 86.59 $\pm$ 1.13 | 82.85 $\pm$ 0.44 | 43.70 $\pm$ 0.83 | 84.18 $\pm$ 0.23 | 86.50 $\pm$ 0.33 | 85.01 $\pm$ 1.46 | 38.24 $\pm$ 0.55 | 85.37 $\pm$ 0.29 | | FedIIR | 68.26 $\pm$ 0.66 | 66.12 $\pm$ 0.74 | 79.48 $\pm$ 0.99 | 63.31 $\pm$ 1.38 | 86.75 $\pm$ 0.98 | 84.75 $\pm$ 1.92 | 40.91 $\pm$ 0.64 | 84.94 $\pm$ 0.49 | 87.77 $\pm$ 0.66 | 86.10 $\pm$ 0.95 | 34.69 $\pm$ 1.07 | 87.66 $\pm$ 0.47 | | FedAvg+FOOGD | 75.09 $\pm$ 0.79 | 73.71 $\pm$ 0.93 | 35.32 $\pm$ 1.02 | 91.21 $\pm$ 0.78 | 88.36 $\pm$ 0.43 | 87.26 $\pm$ 0.86 | 17.78 $\pm$ 0.62 | 96.53 $\pm$ 0.18 | 88.90 $\pm$ 0.29 | 88.25 $\pm$ 0.12 | 12.02 $\pm$ 0.34 | 97.77 $\pm$ 0.41 | | FedRoD | 91.15 $\pm$ 0.87 | 89.90 $\pm$ 0.85 | 47.97 $\pm$ 1.88 | 80.96 $\pm$ 0.90 | 89.62 $\pm$ 0.55 | 87.70 $\pm$ 0.80 | 37.03 $\pm$ 1.40 | 86.50 $\pm$ 0.97 | 87.69 $\pm$ 0.88 | 86.26 $\pm$ 1.19 | 36.13 $\pm$ 1.12 | 86.65 $\pm$ 0.36 | | FOSTER | 90.22 $\pm$ 0.88 | 88.70 $\pm$ 0.82 | 47.40 $\pm$ 1.27 | 77.44 $\pm$ 0.93 | 86.92 $\pm$ 1.85 | 85.82 $\pm$ 1.10 | 42.03 $\pm$ 1.51 | 83.91 $\pm$ 1.11 | 87.83 $\pm$ 1.38 | 85.96 $\pm$ 1.02 | 36.42 $\pm$ 1.14 | 86.19 $\pm$ 0.87 | | FedTHE | 91.05 $\pm$ 0.66 | 89.77 $\pm$ 0.91 | 58.14 $\pm$ 2.79 | 82.04 $\pm$ 1.15 | 89.14 $\pm$ 0.93 | 87.68 $\pm$ 0.41 | 40.28 $\pm$ 2.43 | 85.30 $\pm$ 1.91 | 88.14 $\pm$ 0.24 | 86.18 $\pm$ 0.57 | 35.35 $\pm$ 1.94 | 86.79 $\pm$ 0.37 | | FedICON | 89.06 $\pm$ 0.43 | 89.18 $\pm$ 0.81 | 48.22 $\pm$ 1.48 | 81.28 $\pm$ 0.44 | 75.83 $\pm$ 1.07 | 75.35 $\pm$ 0.36 | 56.19 $\pm$ 1.58 | 79.88 $\pm$ 0.51 | 87.20 $\pm$ 1.13 | 85.39 $\pm$ 0.99 | 35.63 $\pm$ 1.16 | 86.45 $\pm$ 0.41 | | FedRoD+FOOGD | 93.51 $\pm$ 0.65 | 92.74 $\pm$ 0.46 | 32.99 $\pm$ 1.30 | 91.76 $\pm$ 0.26 | 90.46 $\pm$ 0.78 | 90.16 $\pm$ 0.51 | 25.51 $\pm$ 1.46 | 94.19 $\pm$ 0.78 | 89.44 $\pm$ 0.88 | 88.62 $\pm$ 0.37 | 18.91 $\pm$ 0.46 | 96.25 $\pm$ 0.22 |

Authorsrebuttal2024-08-10

Response to Reviewer cFix

Table 2: Main results of federated OOD detection and generalization on Cifar100. We report the ACC of brightness as IN-C ACC, the FPR95 and AUROC of LSUN-C as OUT performance. | Non-IID \ Method | $\alpha=0.1$ | | | | $\alpha=0.5$ | | | | $\alpha=5$ | | | | | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | | | ACC-IN $\uparrow$ | ACC-IN-C $\uparrow$ | FPR95 $\downarrow$ | AUROC $\uparrow$ | ACC-IN $\uparrow$ | ACC-IN-C $\uparrow$ | FPR95 $\downarrow$ | AUROC $\uparrow$ | ACC-IN $\uparrow$ | ACC-IN-C $\uparrow$ | FPR95 $\downarrow$ | AUROC $\uparrow$ | | FedAvg | 51.67 $\pm$ 1.37 | 47.54 $\pm$ 0.48 | 78.35 $\pm$ 1.64 | 67.16 $\pm$ 1.17 | 58.28 $\pm$ 0.48 | 54.62 $\pm$ 0.67 | 72.84 $\pm$ 0.81 | 70.86 $\pm$ 1.52 | 61.40 $\pm$ 0.12 | 56.72 $\pm$ 0.17 | 72.68 $\pm$ 0.34 | 70.59 $\pm$ 0.19 | | FedLN | 52.48 $\pm$ 1.41 | 48.15 $\pm$ 1.57 | 66.94 $\pm$ 1.61 | 74.82 $\pm$ 0.50 | 59.39 $\pm$ 0.72 | 53.86 $\pm$ 1.23 | 68.31 $\pm$ 1.24 | 73.41 $\pm$ 0.33 | 61.00 $\pm$ 0.40 | 56.33 $\pm$ 0.82 | 69.18 $\pm$ 0.46 | 75.87 $\pm$ 0.74 | | FedATOL | 43.65 $\pm$ 0.54 | 41.08 $\pm$ 0.60 | 65.26 $\pm$ 0.96 | 81.64 $\pm$ 0.33 | 60.62 $\pm$ 0.61 | 56.63 $\pm$ 0.91 | 70.10 $\pm$ 0.81 | 79.27 $\pm$ 0.61 | 64.16 $\pm$ 0.81 | 63.61 $\pm$ 0.42 | 80.27 $\pm$ 1.61 | 60.51 $\pm$ 1.75 | | FedT3A | 51.67 $\pm$ 1.37 | 51.50 $\pm$ 0.29 | 78.35 $\pm$ 1.64 | 67.16 $\pm$ 1.17 | 58.28 $\pm$ 0.48 | 55.42 $\pm$ 1.63 | 72.84 $\pm$ 1.56 | 70.86 $\pm$ 1.52 | 61.40 $\pm$ 0.12 | 55.51 $\pm$ 0.96 | 72.68 $\pm$ 0.34 | 70.59 $\pm$ 0.19 | | FedIIR | 51.63 $\pm$ 0.61 | 47.88 $\pm$ 1.19 | 81.91 $\pm$ 0.47 | 63.99 $\pm$ 0.53 | 58.66 $\pm$ 0.41 | 55.72 $\pm$ 0.29 | 77.62 $\pm$ 1.10 | 65.87 $\pm$ 0.46 | 61.70 $\pm$ 0.76 | 57.65 $\pm$ 0.80 | 72.57 $\pm$ 0.37 | 69.07 $\pm$ 0.52 | | FedAvg+FOOGD | 53.84 $\pm$ 0.83 | 51.69 $\pm$ 0.32 | 36.40 $\pm$ 1.11 | 91.41 $\pm$ 0.36 | 61.82 $\pm$ 0.20 | 59.91 $\pm$ 0.31 | 55.70 $\pm$ 0.78 | 86.42 $\pm$ 0.24 | 64.96 $\pm$ 0.51 | 64.18 $\pm$ 0.31 | 57.70 $\pm$ 0.87 | 84.03 $\pm$ 0.15 | | FedRoD | 73.13 $\pm$ 0.85|69.26 $\pm$ 0.41| 66.34 $\pm$ 1.53 | 73.02 $\pm$ 1.82 | 66.88 $\pm$ 0.61 | 61.28 $\pm$ 0.98 | 70.13 $\pm$ 0.86 | 69.48 $\pm$ 0.65 | 61.34 $\pm$ 0.78 | 55.80 $\pm$ 1.21 | 74.86 $\pm$ 0.98 | 67.76 $\pm$ 1.31 | | FOSTER | 72.54 $\pm$ 1.51 | 67.50 $\pm$ 0.57| 61.25 $\pm$ 1.05 | 75.44 $\pm$ 0.89 | 62.45 $\pm$ 0.55 | 57.62 $\pm$ 0.87 | 73.26 $\pm$ 1.13 | 68.71 $\pm$ 0.85 | 53.80 $\pm$ 0.31 | 49.28 $\pm$ 0.74 | 76.94 $\pm$ 1.62 | 65.47 $\pm$ 1.72 | | FedTHE | 73.83 $\pm$ 0.48 | 69.09 $\pm$ 0.56 | 64.73 $\pm$ 0.79 | 75.16 $\pm$ 0.34 | 66.22 $\pm$ 0.68 |61.19 $\pm$ 0.92 | 72.95 $\pm$ 1.84 | 69.38 $\pm$ 1.64 | 61.03 $\pm$ 0.22 | 57.03 $\pm$ 0.16 | 71.43 $\pm$ 0.64 | 69.01 $\pm$ 0.87 | | FedICON | 72.22 $\pm$ 0.72 |67.79 $\pm$ 0.31 | 61.36 $\pm$ 0.39 | 77.12 $\pm$ 0.55 | 65.86 $\pm$ 0.81 |61.83 $\pm$ 0.55 | 69.99 $\pm$ 1.02 | 71.03 $\pm$ 0.39 | 62.11 $\pm$ 0.74|57.62 $\pm$ 0.28 | 70.91 $\pm$ 0.97 | 70.84 $\pm$ 0.73 | | FedRoD+FOOGD | 77.88 $\pm$ 0.28 |75.70 $\pm$ 0.26|58.81 $\pm$ 0.48 | 86.07 $\pm$ 0.39 | 70.30 $\pm$ 0.46 | 68.23 $\pm$ 0.25| 45.19 $\pm$ 0.67 | 89.59 $\pm$ 0.28 | 64.94 $\pm$ 0.79 | 62.56 $\pm$ 0.72 | 65.18 $\pm$ 1.19 |80.47 $\pm$ 0.32| **As we can see, the model performance variance of FOOGD is very small, indicating the performance stability of FOOGD in OOD data. This also validates that our model is effective and practical for federated learning with wild data.** **For weakness 2, we reply that ACC-IN and ACC-IN-C are commonly used accuracies for the two test sets, i.e., in-distribution data (Non-IID data) and covariate-shift data (OOD Generalization), respectively**. We have introduced them in lines 278-2799 of the main paper. This is used in federated learning considering semantic shifts, e.g., FedTHE, where ACC-IN corresponds to original local test, and ACC-IN-C corresponds to corrupted local test. **For weakness 3, the related work, e.g., FedAvg as well as their implementations are listed in lines 664-693, please kindly refer to this part for more details.** In this work, we mainly focus on tackle the OOD shifts in federated learning, which is orthogonal to the existing work that tackles heterogeneity. We will add a section for introducing work related on FL with non-IID data. **The reference you indicated, i.e., ref.46, is the SVHN dataset that is commonly used in detecting OUT samples.** SVHN is a real-world image dataset from Google Street View house numbers, comprising 73,257 training samples and 26,032 testing samples across 10 classes. We introduce it as an OOD detection task in line 267 and line 646 of paper.

Authorsrebuttal2024-08-10

Response to Reviewer cFix

**For weakness 4, we release our project in an anonymous page: https://anonymous.4open.science/r/FOOGD/, and we have released the test model trained on Cifar10 $\alpha=0.5$ for FedAvg-FOOGD, in our supplements, for validating the effectiveness of FOOGD.** **For the response of question 1, please point out your concerns directly. We are willing to clarify all the misunderstandings.** In summary, we believe that some previous rebuttals might have caused a misunderstanding. We hope you can reconsider our response to see if it resolves your issue. If you have any other questions, please kindly consider responding; we would be happy to discuss them. Please provide us your additional concens as soon as possible, since the the rebuttal time is approaching the end.

Reviewer cFix2024-08-11

Thank you for the detailed response. Some concerns have been addressed in the response, so I have raised the mark to the positive rating.

Authorsrebuttal2024-08-11

Response to reviwer cFix

Thank you for your timely response and insightful reviwes. We are delighted to have addressed your concens and to know that you appreciate our work! The conversation with you are greatly beneficial for our work, such as providing mor detailed explainations. Once again, we are sincerely gratitude for your thorough and patient reviews.

Reviewer MUVw7/10 · confidence 4/52024-07-09

Summary

This paper addresses various OOD (Out-of-Distribution) shifts that may occur in federated settings by proposing a unified framework to simultaneously tackle OOD generalization and detection issues. Specifically, this paper introduces FOOGD, which estimates arbitrary client probability densities to create a reliable global distribution that guides both OOD generalization and detection. Besides, it presents $SM^3D$ and SAG to respectively enhance OOD detection and generalization capabilities. Finally, extensive experiments demonstrate the effectiveness of the proposed methods

Strengths

1. This paper proposes a novel unified framework to simultaneously address OOD (Out-of-Distribution) generalization and detection problems in federated settings. 2. The method has theoretical guarantees and empirical analysis to ensure its effectiveness. 3. The structure of this work is logical, and the writing is well-organized.

Weaknesses

1. This paper proposes the FOOGD, $SM^3D$, and SAG methods to address OOD generalization and detection problems. It is essential to clearly define the relationships among these three methods in the abstract and introduction to enhance the readability of the entire paper. 2. This paper uses 'semantic-shift' to describe the issue of unknown class bias in OOD detection, following the traditional machine learning concept[1]. However, in federated learning, this problem is generally referred to as 'concept shift'[2][3]. It is recommended to revise this terminology to align better with the conventions of the federated learning community. 3. In the related work section (Section 2.3), it is recommended to include a discussion on 'concept shift' in federated learning and cite relevant literature. 4. In the methodology section, given the involvement of multiple innovative components, it is recommended to analyze their complexity. References: [1] Haoyue Bai et al. Feed two birds with one scone: Exploiting wild data for both out-of-distribution generalization and detection. ICML, 2023. [2] Panchal K et al. Flash: concept drift adaptation in federated learning. ICML, 2023. [3] Jothimurugesan E et al. Federated learning under distributed concept drift. ICAI, 2023.

Questions

See weaknesses.

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

See weaknesses.

Reviewer RpTd8/10 · confidence 4/52024-07-13

Summary

This paper focuses on the federated learning setup that non-IID, semantic-shift, and covariate-shift take place in the same time. And the authors propose a framework of FOOGD with a unnormalized distribution estimation method, i.e., SM3D, to release the distribution assumption and constraints in heterogeneous modeling. Both federated OOD generalization and detection can be soundly solved via SM3D and SAG modules. Compared with existing methods, FOOGD is validated via extensive experiments for effectiveness. After rebuttal, I raise the Rating from 7 to 8.

Strengths

1. The presentation is straightforward and clear, with detailed and instructive figures. The formula derivation is thorough and essential. 2. The paper addresses a significant challenge in federated learning, focusing on learning from large-scale non-IID data with out-of-distribution (OOD) shifts. 3. The authors conduct extensive and robust experiments, effectively demonstrating the effectiveness of the proposed FOOGD method. 4. Detailed theoretical foundations are laid out for both the estimation model and the consistency of convergence optimization. 5. Utilizing score modeling for federated learning is quite novel and insightful for heterogeneously distributed data.

Weaknesses

1. How to understand IsOUT() function with negative threshold in Equation 9? 2. Score model brings additional computation cost. 3. The correspondence between federated loss of $\mathcal{L}_k$ and the practical loss $\mathcal{L}_s$ and $\mathcal{L}_f$ are not explicitly explained.

Questions

1. How to enhance the generalization for unseen client distribution?

Rating

8

Confidence

4

Soundness

4

Presentation

4

Contribution

3

Limitations

Yes

Reviewer RpTd2024-08-08

Thanks for your response

Thank you for the detailed response, the author effectively addressed my issue, I will increase my grade. Hope you can add these details to your appendix.

Authorsrebuttal2024-08-11

Response to reviewer RpTd

Thank you for appreciating our work. We will make sure to include the relevant details and the addressed specific issues in the camera-ready version.

Reviewer cFix2024-08-10

Thank you for the response. However, I am unhappy with the response since it did not address my concerns completely. All the problems are planned to be addressed in future work rather than the current version, and the authors did not provide evidence to address the issues in the response. Thus, I will make a weak reject decision for the paper.

Authorsrebuttal2024-08-11

Response to Reviwer dEK5

Thank you very much for your reviews and response. We sincerely appreciate it for your time and effort in evaluating our work. We are grateful for your valuable feedback and are particularly thankful for your recognition of the contributions we have made.

Authorsrebuttal2024-08-11

Response to Reviewer MUVw

We are pleased that our work has received professional evaluation and recognition. We are more than willing to incorporate the suggested revisions into our camera-ready version to address the issues raised.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC