pith. sign in

arxiv: 2405.10271 · v5 · submitted 2024-05-16 · 💻 cs.LG · cs.AI· cs.DC· cs.ET

Pruning Federated Models through Loss Landscape Analysis and Client Agreement Scoring

Christian Intern\`o , Elena Raponi , Markus Olhofer , Ali Raza , Thomas B\"ack , Niki van Stein , Yaochu Jin , Barbara Hammer This is my paper

Pith reviewed 2026-05-24 00:38 UTC · model grok-4.3

classification 💻 cs.LG cs.AIcs.DCcs.ET
keywords federated learningmodel pruningloss landscapenon-IID dataclient agreementadaptive pruningresource efficiency
0
0 comments X

The pith

AutoFLIP prunes federated models by mapping the collective loss landscape through one-time client collaboration and then refining sub-networks via ongoing client agreement scoring.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that client data heterogeneity in federated learning can be treated as a resource for revealing the essential structure of the loss landscape rather than a source of instability. It shows that a single shared exploration phase lets clients jointly build this map, after which an adaptive pruning process guided by client agreement identifies efficient sub-networks. If correct, the method delivers lower computational and communication costs while matching or exceeding accuracy in non-IID regimes. The central shift is from mitigating diversity to harnessing it for pruning decisions made before and during training.

Core claim

AutoFLIP performs a one-time federated loss exploration in which clients collaboratively construct a map of the collective loss landscape; this map then directs an adaptive pruning strategy that is continuously adjusted according to client agreement scores throughout subsequent training, thereby locating robust sub-networks from the start.

What carries the argument

The AutoFLIP framework, whose core mechanism is the initial federated loss-landscape mapping followed by client-agreement-guided pruning.

If this is right

  • Computational overhead drops by an average of 52 percent across tested settings.
  • Communication costs fall by more than 65 percent while accuracy remains at state-of-the-art levels.
  • Resource-constrained devices can host larger models without proportional increases in training expense.
  • The same loss-landscape map serves as a stable reference for pruning choices made at any later training stage.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same loss-exploration step could be reused to initialize other federated compression techniques beyond magnitude-based pruning.
  • Client agreement scores might serve as a lightweight diagnostic for detecting when the global model has overfit to particular client subsets.
  • If the loss map proves stable, periodic re-exploration could be omitted, further lowering the upfront cost.

Load-bearing premise

A single one-time federated loss exploration performed before main training produces a sufficiently accurate and stable map of the collective loss landscape to support effective pruning decisions across heterogeneous clients.

What would settle it

An experiment in which pruning decisions derived from the initial loss map produce measurably lower final accuracy than an unpruned baseline under the same non-IID client partitions would falsify the central claim.

Figures

Figures reproduced from arXiv: 2405.10271 by Ali Raza, Barbara Hammer, Christian Intern\`o, Elena Raponi, Markus Olhofer, Niki van Stein, Thomas B\"ack, Yaochu Jin.

Figure 1
Figure 1. Figure 1: Conceptual overview of the AutoFLIP framework. Phase 1 (One-Time): An initial Federated Loss Exploration, where clients provide loss landscape deviations to construct the Global Guidance Matrix. Phase 2 (Continuous): A Dynamic Adaptation phase, where the guidance matrix is continuously refined using participant client agreement scores. The evolved matrix is then used to generate an adaptive pruning mask. t… view at source ↗
Figure 2
Figure 2. Figure 2: FL optimization process. At each communication [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Variance of weight updates (σ 2 ∆W ) over FL rounds for AutoFLIP and FedAvg, illustrating the stabilizing effect of guided pruning. This data-driven trust mechanism ensures the guidance matrix evolves robustly, balancing stability with adaptability to the changing model state. Takeaway 1: Novel Mechanism Our pruning is not random. Instead, it is guided by two signals: a collaborative map of the loss landsc… view at source ↗
Figure 4
Figure 4. Figure 4: Distribution of Gglobal. The absolute frequency (in log-scale) is shown for each normalized deviation. Higher frequencies are recorded for smaller deviation values, indicating that many parameters are irrelevant for loss improvement [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The AutoFLIP hybrid pruning mechanism. The global pruning mask simultaneously removes individual weights (un￾structured, dotted lines) and, consequently, entire neurons when all their connections are removed (structured, high￾lighted nodes). or a convolutional filter), the entire unit is deactivated. This removal of entire components from the network graph is the key to computational savings. The final dis… view at source ↗
Figure 6
Figure 6. Figure 6: Synthesis of the AutoFLIP pruning process for the Six-layer CNN on MNIST. The plot shows the relationship between a parameter’s initial importance (x-axis) and its final importance after adaptation via client agreement (y-axis). The Tp threshold acts as the final decision boundary, pruning all parameters in the shaded region. This visualizes the two primary reasons for pruning: low initial importance or lo… view at source ↗
Figure 7
Figure 7. Figure 7: Global accuracy convergence of AutoFLIP against FedAvg [17], PruneFL [18], EFLPrune [19] and FedMask [20]. TABLE III: Final global model accuracy comparison across all datasets and models. Results are presented as mean ± standard deviation. Best results are highlighted. Algorithm MNIST - CNN CIFAR10 - EffNetB3 Tiny ImageNet - MobileNetV2 CIFAR100 - ResNet FEMNIST - CNN Shakespeare - LSTM FedAvg 0.803 ± 0.0… view at source ↗
Figure 8
Figure 8. Figure 8: Accuracy vs. Efficiency Trade-off. The plot visualizes [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Robustness Sensitivity Analysis. We sweep (a, b) the noise magnitude (σ) with a 20% noisy client scope, and (c, d) the client scope (%) with a 0.1 σ. The results show AutoFLIP’s final accuracy (red) degrades gracefully, while FedAvg’s (green) collapses. 0 50 100 150 200 Communication Rounds 0.0 0.2 0.4 0.6 0.8 1.0 Global Accuracy MNIST 0 50 100 150 200 Communication Rounds CIFAR10 AutoFLIP (Noise-Free) Aut… view at source ↗
Figure 10
Figure 10. Figure 10: Robustness to Data-Level (Label) Noise. On 20% of clients, 30% of local data labels were flipped. On MNIST and on CIFAR10, FedAvg (green) suffers severe accuracy degradation from the divergent updates. AutoFLIP (red) identifies these as low-agreement updates and filters them, maintaining high stability. and its final MNIST accuracy collapses to ≈ 0.6. AutoFLIP (red line), however, is highly resilient. Its… view at source ↗
Figure 11
Figure 11. Figure 11: Ablation study on the number of exploration clients [PITH_FULL_IMAGE:figures/full_fig_p010_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Ablation study on the pruning threshold Tp for CIFAR10 under a pathological non-IID data distribution, based on average accuracy (top) and loss (bottom). this approach is both practical and highly effective. We showed that AutoFLIP consistently discovers superior sparse sub￾networks, leading to significant reductions in computational overhead (averaging 52%) and accelerating convergence to target accuraci… view at source ↗
read the original abstract

The practical deployment of Federated Learning (FL) on resource-constrained devices is fundamentally limited by the high cost of training large models and the instability caused by heterogeneous (non-IID) client data. Conventional pruning methods often treat data heterogeneity as a problem to be mitigated. In this work, we introduce a paradigm shift: we reframe client diversity as a feature to be harnessed. We propose AutoFLIP, a framework that begins not with training, but with a one-time federated loss exploration. During this phase, clients collaboratively build a map of the collective loss landscape, using their diverse data to reveal the problem's essential structure. This shared intelligence then guides an adaptive pruning strategy that is dynamically refined by client agreement throughout training. This approach allows AutoFLIP to identify robust and efficient sub-networks from the outset. Our extensive experiments show that AutoFLIP reduces computational overhead by an average of 52% and communication costs by over 65% while simultaneously achieving state-of-the-art accuracy in challenging non-IID settings.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The manuscript proposes AutoFLIP, a federated learning pruning framework that performs a one-time federated loss exploration phase in which clients collaboratively construct a map of the collective loss landscape from their heterogeneous data. This shared map then guides an adaptive pruning strategy that is dynamically refined using client agreement scores throughout training. The authors claim that the method reduces computational overhead by an average of 52% and communication costs by over 65% while attaining state-of-the-art accuracy in challenging non-IID settings.

Significance. If the central claims hold, the work would be significant because it reframes client data heterogeneity as a source of useful structural information for pruning rather than solely a source of instability. The combination of loss-landscape exploration and client-agreement scoring offers a concrete mechanism for identifying robust sub-networks early, which could improve the practicality of large-model FL on edge devices.

major comments (1)
  1. [Abstract (paragraph describing the framework)] Abstract (paragraph describing the framework): The central premise that a single one-time federated loss exploration phase performed before main training yields a sufficiently accurate and stable collective loss landscape for guiding pruning decisions is load-bearing for the efficiency and accuracy claims. No quantitative evidence (landscape similarity metrics across training rounds, ablation on re-exploration frequency, or analysis of Hessian/gradient drift under non-IID conditions) is supplied to show that the initial map remains effective once local models begin to diverge. This directly affects whether the reported 52% compute and 65% communication reductions can be expected to persist.
minor comments (2)
  1. The abstract asserts quantitative gains and SOTA accuracy without describing the experimental protocol, baseline methods, number of runs, statistical tests, or error bars, making it impossible to judge whether the data support the claims.
  2. Clarify the precise definition and computation of the client agreement score and how it is used to update the pruning mask at each round.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback on the stability of the one-time federated loss exploration phase. We address the major comment below.

read point-by-point responses

  1. Referee: [Abstract (paragraph describing the framework)] Abstract (paragraph describing the framework): The central premise that a single one-time federated loss exploration phase performed before main training yields a sufficiently accurate and stable collective loss landscape for guiding pruning decisions is load-bearing for the efficiency and accuracy claims. No quantitative evidence (landscape similarity metrics across training rounds, ablation on re-exploration frequency, or analysis of Hessian/gradient drift under non-IID conditions) is supplied to show that the initial map remains effective once local models begin to diverge. This directly affects whether the reported 52% compute and 65% communication reductions can be expected to persist.

    Authors: We appreciate the referee identifying this as a load-bearing assumption. The manuscript's empirical results demonstrate that the reported efficiency gains (52% compute, 65% communication) and SOTA non-IID accuracy are achieved and maintained with the one-time exploration, as shown by the end-to-end training curves and final metrics across multiple datasets and non-IID partitions. These outcomes indicate that the initial collective map remains effective in practice. However, we agree that explicit quantitative validation of landscape stability would strengthen the claims. In the revised manuscript we will add (i) cosine similarity metrics between the initial loss landscape and gradients at later rounds, (ii) an ablation varying re-exploration frequency, and (iii) a brief analysis of gradient drift under the non-IID regimes studied. These additions will directly quantify how long the initial map remains useful. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical framework with no load-bearing derivations

full rationale

The paper introduces AutoFLIP as an empirical pruning framework based on a one-time federated loss exploration phase followed by client-agreement-guided adaptation, with performance claims resting on experimental results rather than any mathematical derivation chain. No equations, fitted parameters renamed as predictions, self-definitional constructs, or self-citation load-bearing steps appear in the abstract or described methodology. The approach is self-contained as a practical method whose validity is asserted via external benchmarks and experiments, not by reducing outputs to inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no explicit free parameters, axioms, or invented entities; the central claim rests on the unstated premise that the one-time loss exploration is both cheap and informative enough to replace conventional pruning pipelines.

pith-pipeline@v0.9.0 · 5738 in / 1253 out tokens · 24059 ms · 2026-05-24T00:38:42.114211+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

40 extracted references · 40 canonical work pages · 1 internal anchor

  1. [1]

    The roadmap to 6g: Ai empowered wireless networks,

    K. B. Letaief, W. Chen, Y . Shi, J. Zhang, and Y .-J. A. Zhang, “The roadmap to 6g: Ai empowered wireless networks,”IEEE Communica- tions Magazine, vol. 57, no. 8, pp. 84–90, 2019

    work page 2019

  2. [2]

    A vision of 6g wireless systems: Applications, trends, technologies, and open research problems,

    W. Saad, M. Bennis, and M. Chen, “A vision of 6g wireless systems: Applications, trends, technologies, and open research problems,”IEEE Network, vol. 34, no. 3, pp. 134–142, 2020

    work page 2020

  3. [3]

    Federated learning on non-iid data: A survey,

    H. Zhu, J. Xu, S. Liu, and Y . Jin, “Federated learning on non-iid data: A survey,”Neurocomput., vol. 465, no. C, p. 371–390, Nov. 2021. [Online]. Available: https://doi.org/10.1016/j.neucom.2021.07.098

    work page doi:10.1016/j.neucom.2021.07.098 2021

  4. [4]

    The future of digital health with federated learning,

    N. Rieke, J. Hancox, W. Li, F. Milletar `ı, H. R. Roth, S. Albarqouni, S. Bakas, M. N. Galtier, B. A. Landman, K. Maier-Hein, S. Ourselin, M. Sheller, R. M. Summers, A. Trask, D. Xu, M. Baust, and M. J. Cardoso, “The future of digital health with federated learning,”npj Digital Medicine, vol. 3, no. 1, p. 119, Sep 2020

    work page 2020

  5. [5]

    Advances and open problems in federated learning,

    P. e. a. Kairouz, “Advances and open problems in federated learning,” Hanover, MA, USA, p. 1–210, Jun. 2021. [Online]. Available: https://doi.org/10.1561/2200000083

    work page doi:10.1561/2200000083 2021

  6. [6]

    A survey on edge intelligence and lightweight machine learning support for future applications and services,

    K. Hoffpauir, J. Simmons, N. Schmidt, R. Pittala, I. Briggs, S. Makani, and Y . Jararweh, “A survey on edge intelligence and lightweight machine learning support for future applications and services,”J. Data and Information Quality, vol. 15, no. 2, Jun. 2023. [Online]. Available: https://doi.org/10.1145/3581759

    work page doi:10.1145/3581759 2023

  7. [7]

    A survey on intent-based networking,

    A. Leivadeas and M. Falkner, “A survey on intent-based networking,” Commun. Surveys Tuts., vol. 25, no. 1, p. 625–655, Jan. 2023. [Online]. Available: https://doi.org/10.1109/COMST.2022.3215919

    work page doi:10.1109/comst.2022.3215919 2023

  8. [8]

    Toward intelligent intent-based network slicing for iot systems: Enabling technologies, challenges, and vision,

    D. Haj Hussein and M. Ibnkahla, “Toward intelligent intent-based network slicing for iot systems: Enabling technologies, challenges, and vision,”IEEE Transactions on Network and Service Management, vol. 22, no. 4, pp. 3480–3495, 2025

    work page 2025

  9. [9]

    Skeletonization: A technique for trimming the fat from a network via relevance assessment,

    M. C. Mozer and P. Smolensky, “Skeletonization: A technique for trimming the fat from a network via relevance assessment,” inAdvances in Neural Information Processing Systems, D. Touretzky, Ed., vol. 1. Morgan-Kaufmann, 1988

    work page 1988

  10. [10]

    Optimal brain damage,

    Y . LeCun, J. Denker, and S. Solla, “Optimal brain damage,” inAdvances in Neural Information Processing Systems, D. Touretzky, Ed., vol. 2. Morgan-Kaufmann, 1989. [Online]. Available: https://proceedings.neurips.cc/paper files/paper/ 1989/file/6c9882bbac1c7093bd25041881277658-Paper.pdf

    work page 1989

  11. [11]

    Deep compression: Compressing deep neural network with pruning, trained quantization and huffman coding,

    S. Han, H. Mao, and W. J. Dally, “Deep compression: Compressing deep neural network with pruning, trained quantization and huffman coding,”

    work page

  12. [12]

    Available: https://api.semanticscholar.org/CorpusID: 2134321

    [Online]. Available: https://api.semanticscholar.org/CorpusID: 2134321

    work page

  13. [13]

    Pruning convolutional neural networks for resource efficient inference,

    P. Molchanov, S. Tyree, T. Karras, T. Aila, and J. Kautz, “Pruning convolutional neural networks for resource efficient inference,” in International Conference on Learning Representations, 2017. [Online]. Available: https://openreview.net/forum?id=SJGCiw5gl

    work page 2017

  14. [14]

    Gate decorator: Global filter pruning method for accelerating deep convolutional neural networks,

    Z. You, K. Yan, J. Ye, M. Ma, and P. Wang, “Gate decorator: Global filter pruning method for accelerating deep convolutional neural networks,” inAdvances in Neural Information Processing Systems, H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alch ´e-Buc, E. Fox, and R. Garnett, Eds., vol. 32. Curran Associates, Inc.,

    work page

  15. [15]

    Available: https://proceedings.neurips.cc/paper files/ paper/2019/file/b51a15f382ac914391a58850ab343b00-Paper.pdf

    [Online]. Available: https://proceedings.neurips.cc/paper files/ paper/2019/file/b51a15f382ac914391a58850ab343b00-Paper.pdf

    work page 2019

  16. [16]

    Hrank: Filter pruning using high-rank feature map,

    M. Lin, R. Ji, Y . Wang, Y . Zhang, B. Zhang, Y . Tian, and L. Shao, “Hrank: Filter pruning using high-rank feature map,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020

    work page 2020

  17. [17]

    SNIP: SINGLE-SHOT NETWORK PRUNING BASED ON CONNECTION SENSITIVITY ,

    N. Lee, T. Ajanthan, and P. Torr, “SNIP: SINGLE-SHOT NETWORK PRUNING BASED ON CONNECTION SENSITIVITY ,” inInterna- tional Conference on Learning Representations, 2019

    work page 2019

  18. [18]

    Rigging the lottery: Making all tickets winners,

    U. Evci, T. Gale, J. Menick, P. S. C. Rivadeneira, and E. Elsen, “Rigging the lottery: Making all tickets winners,” inInternational Conference of Machine Learning, 2020

    work page 2020

  19. [19]

    Communication-efficient learning of deep networks from decentralized data,

    H. B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communication-efficient learning of deep networks from decentralized data,” 2023

    work page 2023

  20. [20]

    Model Pruning Enables Efficient Federated Learning on Edge Devices,

    Y . Jiang, S. Wang, V . Valls, B. J. Ko, W.-H. Lee, K. K. Leung, and L. Tassiulas, “Model Pruning Enables Efficient Federated Learning on Edge Devices,”IEEE Transactions on Neural Networks and Learning Systems, vol. 34, no. 12, pp. 10 374–10 386, Dec. 2023

    work page 2023

  21. [21]

    Efficient federated learning on resource- constrained edge devices based on model pruning,

    T. Wu, C. Song, and P. Zeng, “Efficient federated learning on resource- constrained edge devices based on model pruning,”Complex & Intelli- gent Systems, vol. 9, no. 6, pp. 6999–7013, 2023

    work page 2023

  22. [22]

    FedMask: Jointly learning to trust and mask for federated learning,

    G. Gong, X. Liu, Y . Zhang, L. Liu, and Y . Yang, “FedMask: Jointly learning to trust and mask for federated learning,” inProceedings of the 29th ACM International Conference on Multimedia, 2021, pp. 3266– 3274

    work page 2021

  23. [23]

    Importance estimation for neural network pruning,

    P. Molchanov, A. Mallya, S. Tyree, I. Frosio, and J. Kautz, “Importance estimation for neural network pruning,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 11 264–11 272

    work page 2019

  24. [24]

    Global sparse momentum sgd for pruning very deep neural networks,

    X. Ding, g. ding, X. Zhou, Y . Guo, J. Han, and J. Liu, “Global sparse momentum sgd for pruning very deep neural networks,” inAdvances in Neural Information Processing Systems, H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alch ´e-Buc, E. Fox, and R. Garnett, Eds., vol. 32. Curran Associates, Inc., 2019. [Online]. Available: https://proceedings.neurips....

    work page 2019

  25. [25]

    Universal statistics of fisher information in deep neural networks: Mean field approach,

    R. Karakida, S. Akaho, and S.-i. Amari, “Universal statistics of fisher information in deep neural networks: Mean field approach,” inProceedings of the Twenty-Second International Conference on Artificial Intelligence and Statistics, ser. Proceedings of Machine Learning Research, K. Chaudhuri and M. Sugiyama, Eds., vol. 89. PMLR, 16–18 Apr 2019, pp. 1032–...

    work page 2019

  26. [26]

    Federated loss explo- ration for improved convergence on non-iid data,

    C. Intern `o, M. Olhofer, Y . Jin, and B. Hammer, “Federated loss explo- ration for improved convergence on non-iid data,” in2024 International Joint Conference on Neural Networks (IJCNN), 2024, pp. 1–8

    work page 2024

  27. [27]

    Thrun and L

    S. Thrun and L. Pratt,Learning to Learn: Introduction and Overview. Boston, MA: Springer US, 1998, pp. 3–17

    work page 1998

  28. [28]

    Model-agnostic meta-learning for fast adaptation of deep networks,

    C. Finn, P. Abbeel, and S. Levine, “Model-agnostic meta-learning for fast adaptation of deep networks,” inProceedings of the 34th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, D. Precup and Y . W. Teh, Eds., vol. 70. PMLR, 06–11 Aug 2017, pp. 1126–1135

    work page 2017

  29. [29]

    The need for biases in learning generalizations,

    T. M. Mitchell, “The need for biases in learning generalizations,” 2007

    work page 2007

  30. [30]

    Flat minima,

    S. Hochreiter and J. Schmidhuber, “Flat minima,”Neural Comput., vol. 9, no. 1, p. 1–42, Jan. 1997

    work page 1997

  31. [31]

    Guided transfer learning,

    D. Nikoli ´c, D. Andri ´c, and V . Nikoli´c, “Guided transfer learning,” 2023

    work page 2023

  32. [32]

    Using guided transfer learning to predispose ai agent to learn efficiently from small rna-sequencing datasets,

    K. Li, D. Nikoli ´c, V . Nikoli ´c, D. Andri ´c, L. M. Sanders, and S. V . Costes, “Using guided transfer learning to predispose ai agent to learn efficiently from small rna-sequencing datasets,” 2023

    work page 2023

  33. [33]

    Connecting low-loss subspace for personalized federated learning,

    S.-J. Hahn, M. Jeong, and J. Lee, “Connecting low-loss subspace for personalized federated learning,” inProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. ACM, aug 2022

    work page 2022

  34. [34]

    Leaf: A benchmark for federated settings,

    S. Caldas, S. M. K. Duddu, P. Wu, T. Li, J. Kone ˇcn´y, H. B. McMahan, V . Smith, and A. Talwalkar, “Leaf: A benchmark for federated settings,” 2019

    work page 2019

  35. [35]

    Thop: Pytorch-opcounter,

    L. Zhu, “Thop: Pytorch-opcounter,”THOP: PyTorch-OpCounter, 2022

    work page 2022

  36. [36]

    Federated learning with differential privacy: Algorithms and performance analysis,

    K. Wei, J. Li, M. Ding, C. Ma, H. H. Yang, F. Farhad, S. Jin, T. Q. S. Quek, and H. V . Poor, “Federated learning with differential privacy: Algorithms and performance analysis,” 2019

    work page 2019

  37. [37]

    PPFLex: Securing non-IID optimization in federated learning via MPC,

    N. Yuca, C. Intern `o, N. Matyunin, M. Olhofer, B. Hammer, and S. Katzenbeisser, “PPFLex: Securing non-IID optimization in federated learning via MPC,” in1st Workshop on Federated Learning for Critical Applications, 2026. [Online]. Available: https: //openreview.net/forum?id=PTdCBanvv3

    work page 2026

  38. [38]

    On the opportunities and risks of foundation models,

    R. B. et al., “On the opportunities and risks of foundation models,”

    work page

  39. [39]

    On the Opportunities and Risks of Foundation Models

    [Online]. Available: https://arxiv.org/abs/2108.07258

    work page internal anchor Pith review Pith/arXiv arXiv

  40. [40]

    Federatedfactory: Generative one-shot learning for extremely non-iid distributed scenarios,

    A. Moleri, C. Intern `o, A. Raza, M. Olhofer, D. Klindt, F. Stella, and B. Hammer, “Federatedfactory: Generative one-shot learning for extremely non-iid distributed scenarios,” 2026. [Online]. Available: https://arxiv.org/abs/2603.16370

    work page arXiv 2026