CRAFT: Conflict-Resolved Aggregation for Federated Training

2026-05-20 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

CRAFT (Conflict-Resolved Aggregation for Federated Training) is a novel aggregation framework designed to overcome the fundamental bottleneck of conflicting client updates in federated learning (FL) environments with heterogeneous data distributions. This framework redefines the global update as a geometric correction problem, seeking an update that is closest to a reference direction while adhering to conflict-free alignment constraints. CRAFT provides a closed-form expression for this constrained optimization, eliminating the computational overhead typically associated with iterative solvers. Furthermore, it incorporates a layer-wise adaptation mechanism to address conflicts at different feature granularities. Theoretical analysis confirms that CRAFT fosters a common-descent structure and effectively mitigates conflicts through its projection geometry. Extensive experiments on diverse heterogeneous benchmarks demonstrate that CRAFT significantly enhances the global model's accuracy and reduces performance disparities among clients when compared to existing baselines. The source code is publicly available.

Key takeaway

For Machine Learning Engineers deploying federated learning models with heterogeneous data, CRAFT offers a robust solution to improve model performance and fairness. You should consider integrating CRAFT's conflict-resolved aggregation to mitigate performance degradation caused by conflicting client updates. This approach can enhance global model accuracy and significantly reduce performance disparity across your clients, providing a more stable and equitable FL system.

Key insights

CRAFT resolves federated learning update conflicts via geometric projection, improving global accuracy and client fairness.

Principles

• Conflicting client updates degrade federated learning.
• Geometric projection resolves update conflicts.
• Layer-wise adaptation handles feature granularity.

Method

CRAFT formulates aggregation as finding the update closest to a reference direction under conflict-free alignment constraints, solved via a closed-form expression and layer-wise adaptation.

In practice

• Improve global model accuracy.
• Reduce client performance disparity.
• Apply layer-wise conflict adaptation.

Topics

• Federated Learning
• CRAFT Framework
• Model Aggregation
• Heterogeneous Data
• Geometric Optimization
• Client Performance

Code references

• tum-pbs/CRAFT

Best for: Research Scientist, AI Scientist, Machine Learning Engineer

Related on AIssential

• Conflict-Aware Federated Fine-Tuning of Large Language Models with Mixture-of-Experts — FC-MoE resolves expert conflicts in federated MoE LLM fine-tuning, improving stability and performance in non-IID settings.
• Fairness-Aware Federated Learning with Trajectory Shapley Value — TSV and FedTSV provide dynamic, fairness-aware aggregation for federated learning by evaluating client influence on model trajectory.
• Amortizing Federated Adaptation: Hypernetwork Driven LoRA for Personalized Foundation Models — HyperLoRA uses learned operators for federated adaptation, overcoming LoRA's aggregation bias and client initialization lag.
• Giskard : Byzantine Robust and Confidential Aggregation for Large-Scale Decentralized Learning — Giskard enables scalable, confidential, and Byzantine-robust decentralized learning aggregation using a tree-based committee structure and MPC.
• Domain-Adaptive Model Merging Across Disconnected Modes — DMM merges divergent models data-free by synthesizing pseudo-data and distilling knowledge, outperforming existing methods.
• VARS-FL: Validation-Aligned Client Selection for Non-IID Federated Learning in IoT Systems — Validation-aligned, history-aware client selection significantly improves federated learning convergence and stability in non-IID IoT environments.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.