DAFL's distributed augmentation allocation lifts minority-class F1 to 0.183 and shortens convergence on CIFAR-10 federated learning

The paper decomposes the FedAvg convergence bound into a label proportion imbalance term and a conditional gradient heterogeneity term, noting that the former depends only on class proportions and can be directly reduced by locally augmenting under-represented classes, whereas the latter is influenced by augmented-sample quality and is hard to use as an optimization parameter. Existing analyses such as SCAFFOLD and Khaled et al. model data heterogeneity with a single bounded-dissimilarity constant; this work splits heterogeneity into two separately quantifiable sources, making the link between how much to augment and convergence speed explicit. Derived under assumptions A1–A5 (L-smoothness, conditionally unbiased stochastic gradients, bounded class gradients, bounded conditional variance and conditional independence, bounded conditional gradient heterogeneity) via Lemma 1 and Theorem 1, yielding a convergence upper bound with terms including Γ and σ_g²; this is theoretical analysis, and the text does not provide numerical experimental validation of the bound.

The paper formulates distributed augmentation as a latency-aware optimization problem: subject to global class imbalance not exceeding a given threshold, it jointly minimizes augmentation time and estimated training time to find the minimum augmentation for each client-class pair. Prior augmentation methods treat augmentation as a heuristic and can over-augment and prolong training; this work instead seeks a minimum augmentation allocation under a constraint and introduces relative per-sample time costs for augmentation and training per client. Problem 1 is a non-convex integer optimization; the paper gives a polynomial-complexity two-phase greedy algorithm for a suboptimal solution and notes that Phase 1 may overshoot the constraint and Phase 2 cannot correct it; complexity is stated in terms such as O(KC) per iteration.

On six imbalanced configurations of CIFAR-10 with eight classes, eight clients, and full participation, DAFL achieves both higher minority-class F1 and shorter training time on D2, D4, D5, and D6; on D2 minority-class F1 is 0.183±0.055 with convergence time 188.119±86.344 s, versus 0.002±0.003 and 4167.974±3232.481 s for FedProx and 0.023±0.019 and 578.554±271.105 s for Focal Loss. Relative to NIC, WCEL, LB, FedProx, and Focal Loss, DAFL improves minority-class performance and time simultaneously under severe global class imbalance and high label proportion imbalance, whereas LB over-augments and produces duplicate images that overfit when minority classes have very few samples. Results are reported as mean ± standard deviation over three random seeds with a balanced test set; augmentation uses low-complexity rotation and horizontal flipping, the model is a ConvNet with six convolutional and three fully connected layers and 1,343,146 learnable parameters, and aggregation uses FedAvg.

DAFL's benefit depends on the imbalance type: when global class imbalance is low and label proportion imbalance is high it mainly shortens convergence time, when global class imbalance is high and label proportion imbalance is low it improves minority-class F1 at the cost of longer training, and on D6 it trades a slight drop in accuracy and macro F1 for a small minority-class F1 gain. Rather than presenting the method as uniformly dominant, the paper characterizes its applicable regime by the combination of global class imbalance and label proportion imbalance, offering conditional guidance for choosing an augmentation strategy in deployment. Based on comparison tables across the six configurations D1–D6 and curves plotting minority-class F1, accuracy, and macro F1 against training time; D3 is the exception where DAFL's minority-class performance falls below LB, which the paper attributes to over-augmentation for reducing global imbalance causing overfitting.

The framework targets server-coordinated federated learning with full client participation where the server can obtain each client's class counts, suited to settings with constrained client computation and energy where augmentation overhead must be controlled. It decides the minimum augmentation per client-class pair rather than the augmentation method itself, so it can be combined with any class-preserving augmentation technique. The applicable regime reported is: benefits are clearest when both global class imbalance and label proportion imbalance are high; when global class imbalance is low and label proportion imbalance is high it mainly shortens convergence time; when global class imbalance is high and label proportion imbalance is low it trades training time for minority-class F1. Experiments use the first eight classes of CIFAR-10, eight clients, and a balanced test set, with a 1,343,146-parameter ConvNet, FedAvg aggregation, and convergence defined as training-loss difference below 0.001 over five consecutive rounds.

Several open questions remain for a careful reader: the conditional gradient heterogeneity term is influenced by augmented-sample quality and the paper explicitly does not include it in the optimization, so how the choice of augmentation method changes that term is still to be studied; the greedy algorithm only guarantees a suboptimal solution, and Phase 1 may overshoot the global imbalance constraint with no correction in Phase 2, so the actual augmentation amount may exceed what is theoretically needed; several hyperparameters are not tuned and an unknown scaling constant is set to 1, and the text does not explore how these settings affect the conclusions; evaluation is limited to CIFAR-10 with eight classes and eight clients, leaving performance across datasets, client counts, and partial participation unclear; moreover, server-side solve time must be offset by reduced training time, and how this tradeoff shifts across hardware and network conditions still needs measurement.