ICACT20260326 Slide.20
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Slide 20. Appendix continued. Regularization
|
|
ICACT20260326 Slide.19
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Slide 19. Appendix continued. Symmetric Active Negative Loss
|
|
ICACT20260326 Slide.18
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Slide 18. Appendix continued. Method summary.
|
|
ICACT20260326 Slide.17
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Slide 17. Appendix continued. Problem formulation of the work.
|
|
ICACT20260326 Slide.16
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Slide 16. Appendix continued. Contributions of the work
|
|
ICACT20260326 Slide.15
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Slide 15. Appendix. Abstract
|
|
ICACT20260326 Slide.14
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Slide 14. References
Some of the relevant references used in the work.
|
|
ICACT20260326 Slide.13
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Slide 13. Conclusion.
In this work, we introduced FSANL, a single-stage, noise-robust federated learning framework that jointly addresses heterogeneous label noise and non-IID client drift without assuming clean clients.
By unifying symmetric loss, active negative suppression, and global L1 regularization, FSANL achieves stable and robust optimization under severe noise conditions.
Experiments on CIFAR-10 show that FSANL outperforms state-of-the-art methods, particularly when all clients are noisy, and noise intensity is high.
|
|
ICACT20260326 Slide.12
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Slide 12. Results continued.
Figure showing Test accuracy on CIFAR-10 under IID setting with ρ = 1.0, τ = 0.5 across 700 communication rounds.
|
|
ICACT20260326 Slide.11
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Slide 11. Results.
1. Test accuracy (%) on CIFAR-10 under IID setting with varying heterogeneous noise conditions.
2. Test accuracy (%) on CIFAR-10 under non-IID setting with varying heterogeneous noise conditions.
|
|
ICACT20260326 Slide.10
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Slide 10. Experimental settings
1. Dataset and Model
2. Non-IID partition
3. FSANL Hyperparameters
|
|
ICACT20260326 Slide.09
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Slide 9. Method continued.
Overall Algorithm of FSANL workflow.
|
|
ICACT20260326 Slide.08
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Slide 8. Method continued.
Regularization - penalizes deviation from the global model and encourages consistent local updates.
|
|
ICACT20260326 Slide.07
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Slide 7. Method continued.
SANL details
Normalized Cross Entropy (NCE)
Normalized Negative Cross Entropy (NNCE)
Symmetric Cross Entropy (SCE)
|
|
ICACT20260326 Slide.06
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Slide 6. Method.
1. FL objective – learn a global model by minimizing expected loss:
After local training, the server performs model aggregation.
2. FSANL – a single unified loss:
|
|
ICACT20260326 Slide.05
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Slide 5. Related works.
1. Federated learning under Non-IID data
2. Noise-Robust Learning in Federated Settings
3. FSANL unifies robust loss design and client drift mitigation in a single-stage FL framework. Further, no clean-client assumption, no label correction, no auxiliary models.
|
|
ICACT20260326 Slide.04
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Slide 4. Our contributions are summarized as follows:
1. FSANL, a single-stage federated framework that is robust to non-IID distributions and client-level heterogeneous label noise.
2. FSANL introduces a unified loss function combining symmetric and active negative loss components with L1- based global model regularization.
3. Extensive experiments on CIFAR-10 demonstrate the effectiveness and scalability of FSANL compared to state-of-the-art baselines.
|
|
ICACT20260326 Slide.03
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Slide 3. Introduction. Federated Learning enables collaborative model training and provides basic data privacy by aggregating and sending back updated models.
|
|
ICACT20260326 Slide.02
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Slide 2. Outline of the presentation.
|
|
ICACT20260326 Slide.01
[Big slide for presentation]
|
Chrome Text-to-Speach Click!! |
 |
Slide 1. title, authors, and affiliation
|
