Home / Curated / Algorithms / FedDisco (Federated Learning with Discrepancy-Aware Collaboration)

FedDisco (Federated Learning with Discrepancy-Aware Collaboration)

Algorithms AIMech:FedDiscoFederatedLearningWithDiscrepancyAwareCollaboration DRAFT

Aliases: FedDisco, Federated Learning with Discrepancy-Aware Collaboration, Discrepancy-Aware Federated Aggregation

A federated aggregation algorithm for category-imbalanced and non-IID client datasets (Ye et al. 2023, Sections 3.2-5.2, Proposition 4.1, Theorems 4.4-4.5, Equation 7, Algorithm 1) in which each client k computes the L2 discrepancy d_k = ||r_k - u||_2 (or KL/cosine discrepancy) between its local category distribution r_k in Delta^{C-1} and a target global category distribution u (such as the uniform distribution u_c = 1/C) and the server replaces standard dataset-size aggregation weights n_k = |D_k| / sum_m |D_m| with discrepancy-aware collaboration weights p_k = ReLU(n_k - a d_k + b) / sum_{m=1}^K ReLU(n_m - a d_m + b) (Equation 7) when updating w^{(t+1)} = w^{(t)} - eta sum_{k in S_t} p_k Delta_k^{(t)}, where the one-time privacy-preserving auxiliary transmission of D^{aux}_k = R r_k using a random orthogonal matrix R in R^{C x C} (R^T R = I) preserves exact Euclidean norms ||D^{aux}_k - D^{aux}_u||_2 = ||r_k - u||_2 without revealing raw class counts (Section 5.1, Equation 8).

Identity Notes & Concept Scope Distinct from FedNova (AIMech:FedNovaNormalizedAveragingForHeterogeneousFederatedOptimization, Wang et al. 2020), which normalizes client updates by their effective number of local SGD steps tau_eff rather than category-distribution discrepancy d_k = ||r_k - u||_2; and distinct from Agnostic Federated Learning (AIMech:AgnosticFederatedLearning, Mohri et al. 2019), which solves a minimax saddle-point problem max_lambda min_w sum_k lambda_k F_k(w) assigning larger weights to high-loss clients rather than down-weighting category-skewed clients via ReLU(n_k - a d_k + b).

Claims & Grounded Evidence

4 Claims

Mechanism #r469-gradient-diversity-bound-and-convergence-theorem-mechanism

Proves that the weighted gradient-to-global discrepancy satisfies ||sum_{k=1}^K p_k nabla F_k(w) - nabla F(w)|| <= A sum_{k=1}^K p_k d_k under class-wise gradient bound ||nabla F_{k,c}(w)|| <= A (Proposition 4.1, Equation 2) and bounds weighted gradient diversity Lambda(w, p) in Theorem 4.4 (Equation 3), yielding the non-convex convergence upper bound of Theorem 4.5 (Equation 4) whose discrepancy terms T_3, T_5 are strictly reduced by shifting simplex mass from high-d_k to low-d_k clients via p_k propto ReLU(n_k - a d_k + b) (Equation 7 & Appendix C.4).

supports Source: FedDisco: Federated Learning with Discrepancy-Aware Collaboration · Locator: Section 4 Theoretical Analysis (Proposition 4.1, Theorems 4.4-4.5, Equations 2-5), pp. 3-5, and Appendix C.1-C.4, pp. 14-21

Scope #r469-orthogonal-projection-privacy-preserving-discrepancy-scope

Computes client discrepancy scores in a single pre-training step without revealing local class histograms r_k to the server by having clients multiply r_k by a shared random orthogonal matrix R in R^{C x C} (R^T R = I) and upload D^{aux}_k = R r_k, preserving exact L2 discrepancy ||D^{aux}_k - D^{aux}_u||_2 = sqrt((r_k - u)^T R^T R (r_k - u)) = ||r_k - u||_2 (Section 5.1, Equation 8), and combines modularly with FedAvg, FedProx, FedNova, MOON, and FedDyn.

supports Source: FedDisco: Federated Learning with Discrepancy-Aware Collaboration · Locator: Section 5.1 Discrepancy Estimation (Equation 8) & Section 5.2 Algorithm 1, pp. 5-6

Limitation #r469-asymptotic-bias-weight-divergence-tradeoff-limitation

Introduces a non-vanishing O(W_D(n)) asymptotic bias term T_4 = (4 A^2 C / mu) W_D(n)^2 in Theorem 4.5 whenever p != n (W_D(n) = ||p - n||_1 > 0), creating a finite-horizon trade-off between reducing discrepancy drift sum_k p_k d_k and bounding weight divergence ||p - n||_1 controlled by hyperparameters (a, b).

supports Source: FedDisco: Federated Learning with Discrepancy-Aware Collaboration · Locator: Section 4.2 Theorem 4.5 & Practical Guidance (Equations 4-6), pp. 4-5, and Appendix C.4 Equation 54, p. 21

Application #r469-quantity-and-dirichlet-skewed-federated-benchmarks-application

Applied to federated image classification under Dirichlet and pathological category-skewed + quantity-skewed partitions on CIFAR-10, CIFAR-100, TinyImageNet, FMNIST, and CINIC-10, consistently improving FedAvg, FedProx, MOON, and FedDyn.

supports Source: FedDisco: Federated Learning with Discrepancy-Aware Collaboration · Locator: Section 6 Experiments (Tables 1-5, Figures 4-6), pp. 6-9

⚠️ Curated Limitations & Negative Results

Applies specifically to classification tasks with discrete category distributions r_k in Delta^{C-1} and requires choosing hyperparameters (a, b).
Introduces a non-zero weight-divergence bias term W_D(n) = ||p - n||_1 in the asymptotic regime T -> infty.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:FederatedLearning` Federated Learning	broad_match	FedDisco is a federated aggregation algorithm that weights client updates by both dataset size and category-distribution discrepancy to the global target distribution.

Evidence Sources & Primary Literature

3 Sources

Source ID	Publication / Document	Type / Access	Reference & Checksum
`r469-primary-pdf`	FedDisco: Federated Learning with Discrepancy-Aware Collaboration Rui Ye, Mingkai Xu, Jianyu Wang, Chenxin Xu, Siheng Chen, Yanfeng Wang	primary paper full_text	ID: https://proceedings.mlr.press/v202/ye23b.html sha256:88f4ab6a5f98…
`r469-arxiv-abs`	FedDisco: Federated Learning with Discrepancy-Aware Collaboration (arXiv abstract) Rui Ye, Mingkai Xu, Jianyu Wang, Chenxin Xu, Siheng Chen, Yanfeng Wang	primary paper abstract_only	ID: https://arxiv.org/abs/2305.19229 sha256:b23db4f0b605…
`r469-metadata-cr`	FedDisco: Federated Learning with Discrepancy-Aware Collaboration (Crossref metadata) Rui Ye, Mingkai Xu, Jianyu Wang, Chenxin Xu, Siheng Chen, Yanfeng Wang	primary paper abstract_only	ID: https://proceedings.mlr.press/v202/ye23b.html sha256:e10d41fceb7b…

Ontology & Grounding Topology

Open in interactive network graph →

Curation Provenance

Curator: Antigravity

Method: ai_assisted

Model: Gemini

Created: 2026-10-19T09:12:00Z

Updated: 2026-10-19T09:12:00Z

Source File: data/curated/FedDiscoFederatedLearningWithDiscrepancyAwareCollaboration.yaml · Schema: curated_concept.yaml

Explore across X-Mech

Related computational & biological Mechs: Embedding Map Network Graph ProteinTraitsMech (ESM-2 pLM) DUFMech (Domain pLM) TraitMech CultureMech All 14 X-Mech Repositories