Research map / Federated learning and privacy / Personalized and heterogeneous FL

HERO: A Heterogeneity-Aware Benchmark Library for Federated Continual Learning

NeurIPS 2026 · Federated learning and privacy › Personalized and heterogeneous FL · approach: Non-IID federated learning · 0 citations (OpenAlex)

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.

Introduces HERO, a heterogeneity-aware benchmark library for federated continual learning that separates task split, client data split, and client task order, and evaluates 10 FCIL methods on CIFAR-100 and TinyImageNet plus a graph Domain-IL portability study on OGB-MolPCBA.

Problem

Federated continual learning results are hard to compare because papers simultaneously change datasets, task splits, client data splits, task orders, backbones, memory budgets and reporting rules, so robustness to client skew or task-order mismatch cannot be isolated.

Approach

Generate benchmark streams that decouple the task split, the Dirichlet client data split (alpha controls skew) and the client task order (rho controls order mismatch via Kendall distance), reuse the same released stream for every method, and report average accuracy, forgetting and bottom-10% client accuracy together.

Key results

• On CIFAR-100 joint-hard setting, TagFed reaches AFA 42.83 and B10 35.92, while FedCBDR is best in the easy setting with AFA 53.91.
• Under joint-hard heterogeneity TagFed is best on AFA and B10 on both datasets, while FedCBDR leads the easy setting, showing method rankings change across settings.
• On OGB-MolPCBA, AP drops from 25.60 to 22.80 for MOTION as scaffold domains grow from G=5 to G=25, and forgetting rises for all methods.

Surprising findings

• The strongest method under synchronized (easy) evaluation, FedCBDR, is not the most stable under joint heterogeneity, where TagFed leads.
• Average accuracy can hide weak bottom-client performance; FedCBDR has a larger easy-to-joint drop than TagFed and LGA.

Limitations

• HERO-Core covers only image-based FCIL on CIFAR-100 and TinyImageNet; portability and extended tiers are not a dense leaderboard.
• Fixes client progression speed and assumes every client eventually observes every task, so asynchronous clients and missing tasks are not covered.
• Does not provide formal privacy analysis, communication accounting, or energy reporting.

Experimental setup

Datasets: CIFAR-100 TinyImageNet OGB-MolPCBA

Benchmarks: CIFAR-100 TinyImageNet OGB-MolPCBA

Metrics: AFA AF B10 AP

Baselines: Local-Only FedAvg GLFC MFCL TARGET LANDER Re-Fed+ TagFed LGA FedCBDR POWER MOTION

Builds on: GLFC FedAvg LEAF FLamby

Closest papers by meaning

• Resource-Constrained Federated Continual Learning: What Does Matter? (NeurIPS 2025 · similarity 0.89)
• Accurate Forgetting for Heterogeneous Federated Continual Learning (ICLR 2024 · similarity 0.88)
• Handling Spatial-Temporal Data Heterogeneity for Federated Continual Learning via Tail Anchor (CVPR 2025 · similarity 0.88)
• CAN: Leveraging Clients As Navigators for Generative Replay in Federated Continual Learning (ICML 2025 · similarity 0.88)
• Efficient Federated Learning against Heterogeneous and Non-stationary Client Unavailability (NeurIPS 2024 · similarity 0.87)
• Reviving Stale Updates: Data-Free Knowledge Distillation for Asynchronous Federated Learning (NeurIPS 2026 · similarity 0.87)