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

Federated Concept-Based Models: Interpretable models with distributed supervision

NeurIPS 2026 · Federated learning and privacy › Personalized and heterogeneous FL · approach: Federated learning with pretrained models · 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.

Proposes Federated Concept-based Models (F-CMs), which aggregate client concept graphs and adapt concept-based model architectures module-wise as concept sets and dependencies evolve in federated learning.

Problem

Concept-based models need concept annotations that are scarce within single institutions, and standard federated learning assumes a fixed shared architecture, which clashes with CMs whose architecture depends on the evolving concept set and dependencies under heterogeneous, non-stationary participat

Approach

Shared CM is decomposed into concept-specific and task-specific modules; the server aggregates client DAGs by maximum consensus over edge confidences, adds or rewires only affected modules (warm-start), and aggregates each module only from clients that supervise it.

Key results

• F-CMs reach task accuracy close to centralized upper bound across benchmarks, e.g., Asia ~80.1-80.7% vs centralized ~80%, and near 100% concept coverage.
• Static federated variant S-F-CMs and localized training trail F-CMs, e.g., Hailfinder OpaqNN 54.0% (S-F-CMs) vs 68.0% (F-CMs) style gaps reported in the table.
• F-CMs modify only 8%-61% of parameters during adaptation versus 100% for reinitialization baseline.

Surprising findings

• Localized models show higher intervention variance because they may encode information about unobserved concepts into predicted ones.

Limitations

• Existing CM architectures were not designed for evolving concept spaces, so adaptation is not fully efficient.
• Privacy analysis is limited; information leakage not systematically characterized.
• Potential new attack surfaces from concept-based federated models are unexplored.

Experimental setup

Datasets: Asia Sachs Alarm Insurance Hailfinder SIIM-Pneumothorax CheXpert CheXpert Plus

Benchmarks: Asia Sachs Alarm Insurance Hailfinder SIIM-Pneumothorax CheXpert

Metrics: task accuracy concept coverage label accuracy under interventions parameter change fraction

Baselines: OpaqNN CBM CEM CGM C$^2$BM FedCBM FCL S-F-CMs Localized training

Builds on: Concept Bottleneck Models Concept Embedding Models Causally Reliable Concept Bottleneck Models FedAvg Concept Graph Models

Closest papers by meaning

• FedSA: A Unified Representation Learning via Semantic Anchors for Prototype-based Federated Learning (AAAI 2025 · similarity 0.85)
• Federated Learning with Heterogeneous LLMs: Integrating Small Student Client Models with a Large Hungry Model (AAAI 2025 · similarity 0.85)
• Federated Model Heterogeneous Matryoshka Representation Learning (NeurIPS 2024 · similarity 0.85)
• FLUX: Efficient Descriptor-Driven Clustered Federated Learning under Arbitrary Distribution Shifts (NeurIPS 2025 · similarity 0.85)
• FedDAT: An Approach for Foundation Model Finetuning in Multi-Modal Heterogeneous Federated Learning (AAAI 2024 · similarity 0.84)
• FlowerTune: A Cross-Domain Benchmark for Federated Fine-Tuning of Large Language Models (NeurIPS 2025 · similarity 0.84)