CRIP: Channel Level Representation Injection for Personalized One-Shot Federated Learning

One-shot federated learning (OSFL) aims to train collaborative models with only a single round of communication to save bandwidth and preserve privacy. However, when clients have heterogeneous data domains, existing OSFL methods often fail because they rely on parameter-space aggregation, which cannot effectively reconcile the divergent feature distributions caused by domain shifts. This paper asks: can we achieve effective cross-client knowledge transfer in OSFL by operating directly in the representation space rather than the parameter space?

The authors propose Channel Level Representation Injection (CRIP). Instead of aggregating model weights, CRIP treats the feature extractors of other clients as sources of transferable knowledge. The process involves three steps:

1. Feature Extraction: Clients share their feature extractors with a central server, which broadcasts them to all participants.
2. Channel Alignment: Each client uses a small local calibration batch to compute channel-wise similarity scores (using Centered Kernel Alignment) between its own features and those of other clients.
3. Residual Injection: For each input, the target client injects the aligned features from source clients as residuals into its own activation maps. This allows the model to incorporate cross-domain knowledge while keeping its own parameters frozen.

CRIP consistently outperforms both local models and state-of-the-art OSFL baselines across major domain-heterogeneous benchmarks, including DomainNet, PACS, and Office-Home. The authors demonstrate that because parameters become entangled with domain-specific noise during local training, they are poor candidates for aggregation. In contrast, representations retain more transferable semantic structure. By performing alignment at the channel level, CRIP effectively filters out domain-specific noise and injects only the most compatible features, leading to superior personalization.

AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.