Research map / Learning theory and algorithms / Fairness
Aequa: Fair Model Rewards in Collaborative Learning via Slimmable Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Aequa, a collaborative learning framework that trains a slimmable global model and allocates model widths post-training (via simulated annealing on a fairness cost) so that reward model performance tracks client contributions.
Problem
Fair distribution of model rewards in collaborative/federated learning: participants should receive differentiated final models whose accuracy is commensurate with their contributions, without the free-rider problem.
Approach
Train one global slimmable network with random-width sampling so that subnetworks of different widths are nested and performance degrades gracefully with width; then assign widths to clients via a post-training fair allocation (simulated annealing on a mean/variance utility cost), with TEEs on clients to keep intermediate models confidential, and extend to training-time rewards using contribution
Key results
- Aequa achieves the highest mean accuracy in 17 of 24 predictive-performance settings, versus 7/24 for FedAvg, 0/24 for CGSV and IAFL.
- Aequa attains the highest Pearson correlation between final accuracy and standalone accuracy in 24/24 main settings (e.g., 0.97 to 1.00 in homogeneous partitions versus at most 0.31 for baselines in some cases).
- On FEMNIST (3,597 users), Aequa reaches 73.10% accuracy versus 71.48% for FedAvg-FT, with Pearson 0.9888 versus 0.2904.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Aequa's contribution-based slim-model reward allocation instead of FedAvg-FT | FEMNIST · FEMNIST | accuracy: 71.48 -> 73.10; +1.6 pp | Helped |
| Use Aequa's contribution-based slim-model reward allocation instead of FedAvg-FT | FEMNIST · FEMNIST | Pearson correlation: 0.2904 -> 0.9888 | Helped |
| Use Aequa instead of FedAvg | federated classification · 24 dataset-partition settings | balanced accuracy: 17/24 vs 7/24 settings best | Mixed |
| Use Aequa's contribution-based slim-model reward allocation instead of baseline reward methods | federated classification · 24 dataset-partition settings | Pearson correlation: 24/24 cases | Helped |
| Use Aequa instead of FedAvg-FT | FEMNIST · FEMNIST | training time per communication round: 42±2.9 s -> 55±3.1 s; 1.309× (≈30.9%) increase | Hurt |
| Reduce the minimum width p_min to 0.1 (ablation) | Aequa · quantity-skew scenarios | collaboration gain spread performance: 11/12 cases; overall total to 48/54 cases | Helped |
Experimental setup
Datasets: MNIST Fashion-MNIST SVHN CIFAR-10 CIFAR-100 Stanford Sentiment Treebank FEMNIST
Benchmarks: MNIST Fashion-MNIST SVHN CIFAR-10 CIFAR-100 SST FEMNIST
Metrics: balanced accuracy Pearson correlation collaboration gain spread
Baselines: FedAvg FedAvg-FT CGSV IAFL standalone training
Builds on: slimmable networks FedAvg IAFL CGSV simulated annealing
Closest papers by meaning
- A Kernel Perspective on Distillation-based Collaborative Learning (NeurIPS 2024 · similarity 0.85)
- CoRe: Collaborative Reasoning via Cross Teaching (ICML 2026 · similarity 0.85)
- CoBo: Collaborative Learning via Bilevel Optimization (NeurIPS 2024 · similarity 0.84)
- Paid with Models: Optimal Contract Design for Collaborative Machine Learning (AAAI 2025 · similarity 0.84)
- Unextractable Protocol Models: Collaborative Training and Inference without Weight Materialization (NeurIPS 2025 · similarity 0.83)
- Fair Graph U-Net: A Fair Graph Learning Framework Integrating Group and Individual Awareness (AAAI 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.