Organizations: School of Intelligent Science and Engineering / School of Artificial Intelligence Industry, Jinan University Zhuhai Campus, Zhuhai 519070, China · School of Intelligent Civil and Ocean Engineering, Harbin Institute of Technology, Shenzhen, Shenzhen 518055, China
Federated learning combines complementary knowledge across clients to alleviate local data limitations without sharing raw samples. However, data heterogeneity can cause conflicting updates and hinder generalization. Update coordination reduces interference among client updates, while flatness optimization reduces sensitivity to parameter perturbations to improve generalization. Existing joint designs largely target specific algorithms, limiting flexible combinations of these approaches. To address this problem, we propose a unified federated learning framework that bridges update coordination and flatness optimization. Clients separately accumulate descent signals and flatness corrections along the same local trajectory. The server coordinates both groups with independent states before fusing their outputs into a global update. Explicit conversion between trajectory signals and model updates enables different coordination strategies to work with compatible local or globally guided flatness mechanisms without changing the inference architecture. Experiments show that the evaluated combinations generally improve average and worst-client accuracy over update coordination alone, demonstrating the effectiveness of the proposed framework in combining these complementary optimization mechanisms.
Figures & tables
Figure 1: The bridging role of the proposed framework. (a) A coordination strategy A and a flatness mechanism P are selected independently through their adapters. FedSAM and FedASAM denote the local SAM-guided and ASAM-guided federated variants. (b) The server coordinates descent and flatness signals with independent states, fuses the outputs, and applies the shared-model update through U . (c) The framework supports different pairings through the same coordination interface, with FedMDFG and FedGF highlighted as an example. Checks indicate adapter-enabled compositions, and ellipses indicate extensibility beyond the listed examples.
Figure 2: An overview of the proposed framework. Clients form h=p−g from paired gradient evaluations and separately accumulate descent and flatness signals along the same local trajectory. The server coordinates the two groups, fuses dt=dtg+λdth , and updates the shared model. Yellow and red dashed arrows carry client descent and flatness signals, respectively, while black arrows represent fused updates. Purple solid, gray dashed, and blue dash-dotted arrows show forward propagation, backpropagation, and model broadcast. Blue and pink layers denote shared and optional client-specific parameters. With L layers and s shared layers, xℓ denotes a forward feature, δr indexes backward signals from the output ( r=L−ℓ+1 ), and dt,ℓ denotes the layer- ℓ block of dt . The landscape and clothing icons are schematic.
Flatness mechanism
FedFV (IJCAI 2021)
FedMGDA+ (IEEE TNSE 2022)
FedMDFG-Agg (AAAI 2023)
FedLD-Agg (MICCAI 2024)
FedAWA-Agg (CVPR 2025)
Mean ↑
Worst ↑
Gain ↑
Mean ↑
Worst ↑
Gain ↑
Mean ↑
Worst ↑
Gain ↑
Mean ↑
Worst ↑
Gain ↑
Mean ↑
Worst ↑
Gain ↑
FedAvg (AISTATS 2017)
Mean: 92.25±0.24 Worst: 90.10±0.18
Ordinary
95.16±0.29
93.38±0.69
—
94.91±0.14
93.35±0.39
—
95.20±0.11
93.85±0.66
—
95.00±0.10
93.65±0.26
—
95.10±0.12
93.55±0.13
—
+ FedSAM (ICML 2022)
95.62±0.07
94.40±0.30
+0.46±0.31
95.67±0.04
94.40±0.30
+0.76±0.11
95.73±0.11
94.47±0.16
+0.53±0.11
95.70±0.14
94.50±0.18
+0.69±0.09
95.63±0.05
94.27±0.31
+0.53±0.13
+ FedASAM (ECCV 2022)
95.67±0.14
94.25±0.13
+0.51±0.42
95.68±0.08
94.43±0.23
+0.77±0.13
95.73±0.06
94.45±0.39
+0.53±0.10
95.65±0.03
94.52±0.25
+0.65±0.11
95.70±0.11
94.53±0.10
+0.60±0.09
+ FedGF (ICML 2024)
95.59±0.17
94.28±0.23
+0.43±0.31
95.61±0.23
94.08±0.54
+0.70±0.18
95.65±0.10
94.20±0.41
+0.45±0.16
95.58±0.03
94.38±0.26
+0.58±0.07
95.67±0.11
94.28±0.38
+0.57±0.19
Table 2: Fashion-MNIST results across seeds 0, 1, and 2 at validation-selected checkpoints. Entries report mean ± sample standard deviation ( n=3 ) for macro and worst-client test accuracy (%), and paired macro gains Δ (pp) over the corresponding ordinary coordinator. Bold denotes a combination mean no lower than its ordinary baseline before rounding. FedAvg is a shared parameter-averaging reference.
Flatness mechanism
FedFV (IJCAI 2021)
FedMGDA+ (IEEE TNSE 2022)
FedMDFG-Agg (AAAI 2023)
FedLD-Agg (MICCAI 2024)
FedAWA-Agg (CVPR 2025)
Mean ↑
Worst ↑
Gain ↑
Mean ↑
Worst ↑
Gain ↑
Mean ↑
Worst ↑
Gain ↑
Mean ↑
Worst ↑
Gain ↑
Mean ↑
Worst ↑
Gain ↑
FedAvg (AISTATS 2017)
Mean: 28.69±0.73 Worst: 22.61±1.36
Ordinary
52.78±0.83
46.65±1.45
—
52.59±0.62
47.71±1.65
—
53.55±1.32
49.01±0.15
—
52.85±1.15
47.90±1.67
—
52.10±1.20
46.31±1.18
—
+ FedSAM (ICML 2022)
54.63±1.04
50.25±2.77
+1.85±0.47
57.15±0.85
52.47±1.35
+4.57±0.57
53.61±1.47
49.11±1.71
+0.07±0.17
55.24±1.81
50.94±1.54
+2.38±0.67
57.52±0.89
52.83±0.80
+5.42±0.73
+ FedASAM (ECCV 2022)
61.06±0.78
56.59±1.60
+8.28±0.33
61.02±1.21
56.37±1.37
+8.43±0.60
61.25±1.21
56.54±1.38
+7.70±0.48
59.56±1.12
55.33±1.07
+6.71±0.22
59.54±1.10
53.96±1.68
+7.44±0.34
+ FedGF (ICML 2024)
53.34±0.90
48.11±1.63
+0.56±0.45
54.89±0.85
50.46±0.82
+2.30±0.26
53.84±0.53
49.20±0.50
+0.29±1.15
53.01±1.76
48.55±2.21
+0.15±0.61
53.94±1.18
49.05±1.96
+1.84±0.09
Table 3: CIFAR-100 results across seeds 0, 1, and 2 at validation-selected checkpoints. Entries report mean ± sample standard deviation ( n=3 ) for macro and worst-client test accuracy (%), and paired macro gains Δ (pp) over the corresponding ordinary coordinator. Bold denotes a combination mean no lower than its ordinary baseline before rounding. FedAvg is a shared parameter-averaging reference.
Flatness mechanism
FedFV (IJCAI 2021)
FedMGDA+ (IEEE TNSE 2022)
FedMDFG-Agg (AAAI 2023)
FedLD-Agg (MICCAI 2024)
FedAWA-Agg (CVPR 2025)
Mean ↑
Worst ↑
Gain ↑
Mean ↑
Worst ↑
Gain ↑
Mean ↑
Worst ↑
Gain ↑
Mean ↑
Worst ↑
Gain ↑
Mean ↑
Worst ↑
Gain ↑
FedAvg (AISTATS 2017)
Mean: 75.07±0.30 Worst: 65.15±0.83
Ordinary
76.57±0.30
66.53±2.07
—
75.95±0.92
66.60±2.63
—
75.65±0.80
65.50±1.75
—
75.10±0.74
65.64±0.72
—
71.76±0.66
62.54±1.40
—
+ FedSAM (ICML 2022)
79.98±0.27
70.31±0.74
+3.41±0.55
77.77±0.33
67.22±0.55
+1.81±0.73
79.45±0.18
69.21±0.52
+3.80±0.71
75.90±0.90
65.91±1.40
+0.80±0.39
72.61±0.15
63.71±0.90
+0.85±0.79
+ FedASAM (ECCV 2022)
79.97±0.07
69.76±0.72
+3.40±0.28
80.88±0.23
72.03±1.14
+4.93±0.25
79.92±0.39
69.48±0.71
+4.27±0.49
79.40±0.78
70.03±2.00
+4.29±0.72
74.13±0.73
64.26±1.75
+2.37±0.26
+ FedGF (ICML 2024)
80.50±0.49
72.03±0.52
+3.93±0.23
80.22±0.86
70.52±1.24
+4.26±0.47
77.56±0.69
67.84±1.64
+1.91±1.19
79.77±0.14
70.52±1.29
+4.66±0.60
79.94±0.30
71.89±0.93
+8.17±0.76
Table 4: Office-Home test results across seeds 0, 1, and 2. Entries report mean ± sample standard deviation for client-macro and worst-client accuracy (%), and paired macro gains Δ (pp) over the corresponding ordinary coordinator. Bold indicates an accuracy at least as high as the ordinary baseline. FedAvg is a shared reference. FedGloSS uses a localized variant.
Flatness mechanism
FedFV (IJCAI 2021)
FedMGDA+ (IEEE TNSE 2022)
FedMDFG-Agg (AAAI 2023)
FedLD-Agg (MICCAI 2024)
FedAWA-Agg (CVPR 2025)
Mean ↑
Worst ↑
Gain ↑
Mean ↑
Worst ↑
Gain ↑
Mean ↑
Worst ↑
Gain ↑
Mean ↑
Worst ↑
Gain ↑
Mean ↑
Worst ↑
Gain ↑
FedAvg (AISTATS 2017)
Mean: 77.40±0.28 Worst: 68.42±0.43
Ordinary
78.50±1.81
70.29±4.34
—
79.27±2.58
70.68±5.78
—
78.90±0.20
71.06±0.49
—
77.78±1.40
67.71±3.10
—
78.15±1.22
70.14±0.66
—
+ FedSAM (ICML 2022)
78.47±1.32
70.07±1.47
-0.04±2.91
81.83±1.25
75.39±1.23
+2.56±2.48
78.14±1.80
68.46±2.34
-0.76±1.99
76.90±2.62
68.06±3.16
-0.88±1.83
78.66±1.08
70.19±1.51
+0.52±2.28
+ FedASAM (ECCV 2022)
81.44±0.82
74.74±0.39
+2.94±2.30
82.30±1.39
75.88±1.22
+3.03±3.94
80.26±1.52
72.76±2.72
+1.36±1.70
79.43±1.31
72.01±2.01
+1.65±0.10
81.55±1.03
75.33±0.98
+3.41±1.93
+ FedGF (ICML 2024)
80.31±0.65
73.70±0.56
+1.81±2.46
79.68±1.65
70.94±1.08
+0.40±4.00
76.94±1.46
68.85±2.54
-1.96±1.60
80.17±0.60
73.89±0.44
+2.39±0.96
80.74±0.47
74.42±1.06
+2.60±1.68
Table 5: Retina test results at validation-selected checkpoints. Entries report mean ± sample standard deviation for macro and worst-client accuracy (%), and reported macro gains Δ (pp) over the corresponding ordinary coordinator. Bold denotes a combination mean no lower than its ordinary baseline. FedAvg is shown once as a shared reference. FedGloSS uses a localized variant.
Figure 3: Effect of adding FedASAM on CIFAR-100. Points are test-accuracy means over three seeds.
Figure 4: Branch mixing and the server update. We replay nine Fashion-MNIST A3 blocks from the seed-2 checkpoint at epoch 200. (a) uses a common scale for pairwise inner products. (b) compares predicted loss decreases at equal FedFV step lengths in block 8. Direct is 100% per client. Labels give relative changes.
Figure 5: Validation accuracy for seed 2. Rows show Fashion-MNIST and CIFAR-100. Columns show client-macro and worst-client accuracy. FedASAM uses direct coordination of the perturbed gradient. Curves and bands show centered seven-epoch means and ± one sample standard deviation.
Figure 6: Fashion-MNIST diagnostics for seed 2. Top shows the fraction of client pairs with negative descent cosine before aggregation. Lines and bands show centered nine-epoch means and ± one sample standard deviation. Bottom shows the macro validation-loss increase under shared-backbone gradient ascent at the epoch-200 checkpoints. Radii are relative to the shared parameter norm. The loss axis uses a symmetric log scale.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Coordinator
Ordinary
FedASAM (direct)
FedASAM (separate)
FedGF
FedLESAM
FedGloSS
FedFV
95.36 / 93.85
95.62 / 94.10
95.68 / 94.75
95.76 / 94.50
95.27 / 93.95
95.27 / 94.20
FedMGDA+
95.02 / 93.60
95.77 / 94.70
95.70 / 94.25
95.88 / 94.70
94.90 / 93.80
95.22 / 93.50
FedMDFG-Agg
95.32 / 94.10
95.78 / 94.70
95.83 / 94.65
95.59 / 94.10
95.30 / 93.95
95.24 / 93.85
FedAWA-Agg
95.19 / 93.65
95.83 / 94.50
95.58 / 94.20
95.72 / 94.65
95.21 / 93.25
95.36 / 93.70
FedLD-Agg
95.11 / 93.95
95.63 / 94.40
95.85 / 94.45
95.61 / 94.10
94.92 / 93.65
95.14 / 93.50
Appendix
Table S1 : Additional Fashion-MNIST seed-2 detail, supplementing the three-seed summary in the main paper. Each entry gives macro / worst-client test accuracy (%). All methods use the same 200-epoch budget and validation selection. Both FedASAM variants use joint probes with ρ=2 . The global probes use shared coordinates and ρ=0.05 . The FedAvg reference is 92.15 / 90.05.
Coordinator
Mechanism
Macro
Worst
Gain
Epoch
FedFV
Ordinary
53.68
47.16
176
FedFV
SAM
53.00
48.61
-0.68
299
FedFV
FedGF
54.37
46.67
+0.69
255
FedFV
FedGloSS
54.74
50.86
+1.07
276
FedMGDA+
Ordinary
53.27
47.65
147
FedMGDA+
SAM
51.50
46.42
-1.76
295
Appendix
Table S2 : Earlier CIFAR-100 seed-2 reference runs, retained separately from the updated three-seed main results. All combinations use separate coordination and 300 data epochs. Macro and worst-client accuracy are percentages, with gains in pp from unrounded records. The epoch column identifies the validation-selected checkpoint. These runs use SAM with ρ=0.5 and global probes with ρ=0.05 . The FedAvg reference is 29.16 / 23.13, selected at epoch 299.
Sharpness-aware minimization (SAM) is an effective method for improving the generalization of federated learning (FL) by steering local training toward flat minima. Under data heterogeneity, however, device-side SAM searches for locally flat basins that are incompatible with the flat region preferred by the global objective. We identify this structural failure mode as flatness incompatibility, which explains why improving local flatness alone may provide limited training and generalization improvement for the global model. We reveal that flatness incompatibility arises from data heterogeneity and the friendly adversary phenomenon, and is further amplified by local updates and partial device participation. To mitigate this issue, we propose Federated Learning with variance-suppressed sharpness-aware minimization (FedVSSAM), which constructs a variance-suppressed adjusted direction and uses it consistently in local flatness search, local descent, and global update. FedVSSAM anchors both perturbation and update directions to a more stable global direction, instead of correcting only an isolated local perturbation. We establish non-convex convergence guarantees of FedVSSAM and prove that the mean-square deviation between the adjusted direction and the global gradient is effectively controlled. Experiments demonstrate that FedVSSAM mitigates flatness incompatibility and outperforms the baselines across diverse FL settings.
Bingnan Xiao, Yuan Gao, Bingcong Li +3
Fudan University · ETH Zurich · Edith Cowan University +1
In federated learning (FL), multi-step local updates and data heterogeneity usually lead to sharper global minima, which degrades the performance of the global model. Popular FL algorithms integrate sharpness-aware minimization (SAM) into local training to address this issue. However, in the high data heterogeneity setting, the flatness in local training does not imply the flatness of the global model. Therefore, minimizing the sharpness of the local loss surfaces on the client data does not enable the effectiveness of SAM in FL to improve the generalization ability of the global model. We define the flatness distance to explain this phenomenon. By rethinking the SAM in FL and theoretically analyzing the flatness distance, we propose a novel FedNSAM algorithm that accelerates the SAM algorithm by introducing global Nesterov momentum into the local update to harmonize the consistency of global and local flatness. FedNSAM uses the global Nesterov momentum as the direction of local estimation of client global perturbations and extrapolation. Theoretically, we prove a tighter convergence bound than FedSAM by Nesterov extrapolation. Empirically, we conduct comprehensive experiments on CNN and Transformer models to verify the superior performance and efficiency of FedNSAM. The code is available at https://github.com/junkangLiu0/FedNSAM.
Junkang Liu, Fanhua Shang, Yuxuan Tian +2
College of Intelligence and Computing, Tianjin University, Tianjin, China · College of Management and Economics, Tianjin University, Tianjin, China · Medical College, Tianjin University; Peng Cheng Laboratory, Tianjin, China +1
The aggregation of conflicting client updates remains a fundamental bottleneck in federated learning (FL) under heterogeneous data distributions. Naive averaging can produce a global update that improves the global objective while conflicting with specific clients, causing degradation for those clients. In this work, we propose CRAFT (Conflict-Resolved Aggregation for Federated Training), a new aggregation framework that treats the global update as a geometric correction problem. We formulate aggregation as finding the update closest to a reference direction while satisfying conflict-free alignment constraints. We derive a closed-form expression for the constrained optimization problem, avoiding the computational overhead of iterative solvers. Furthermore, we use a layer-wise adaptation to address conflicts at varying feature granularities. We provide a theoretical analysis showing that CRAFT promotes a common-descent structure and mitigates conflicts through its projection geometry. Extensive experiments on heterogeneous benchmarks demonstrate that CRAFT improves the accuracy of the global model while reducing performance disparity across clients compared with state-of-the-art baselines. The source code for CRAFT is available at https://github.com/tum-pbs/CRAFT.
Ziqi Wang, Qiang Liu, Nils Thuerey
Department of Mathematics, Friedrich-Alexander-Universität Erlangen-Nürnberg, Erlangen, Germany. · School of Computation, Information and Technology, Technical University of Munich, Garching, Germany. · Munich Center for Machine Learning, Munich, Germany.