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

Gains: Fine-grained Federated Domain Adaptation in Open Set

NeurIPS 2025 · Federated learning and privacy › Personalized and heterogeneous FL · approach: Federated learning with pretrained models · 2 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 Gains, a federated domain adaptation method for open-set settings that discriminates domain- vs class-incremental new clients via encoder feature and classifier parameter changes, uses contribution-driven aggregation, and an anti-forgetting regularizer to preserve source performance.

Problem

Federated learning where new clients with unseen classes or domains join over time; existing methods either handle only class increments (FOSDA) or sacrifice source-domain accuracy and adaptation speed.

Approach

Split the model into encoder and classifier; detect new knowledge from feature-distance changes (Manhattan) on public data and classifier-parameter changes (Euclidean) to tell class from domain increments; then aggregate source clients with contribution weights per component and regularize source clients toward their pre-join models (anti-forgetting).

Key results

• On DigitFive mild shift, Gains reaches 99.34% T-Acc, 93.21% S-Acc and 94.44% G-Acc, versus FedHEAL 22.60/99.29/83.95 and AutoFedGP 68.11/0.00/13.62.
• On DigitFive medium shift, Gains achieves 97.91% T-Acc and 90.09% S-Acc, exceeding FedHEAL's 93.68% T-Acc and 88.71% S-Acc.
• Convergence: Gains converges in 5 rounds (807.45 time units) versus 40 rounds for FedHEAL and 20 for FedAVG; ablation shows removing AFM drops mild-shift source accuracy to 9.24%.

Measured effects reported in the paper

Surprising findings

• Without the anti-forgetting mechanism, mild-shift source accuracy collapses to 9.24%, showing source forgetting is most severe under class increment.
• Several baselines (FOSDA, SemiFDA, AutoFedGP, FedProto) yield near-zero accuracy on one of the target or source metrics in the mild shift setting.

Limitations

• Knowledge discovery thresholds T_F and T_C are set manually.
• Only domain-increment and class-increment cases are considered; task increment and combined increments are not studied.
• Reported results are on small digit and review datasets with simple models (CNN, LSTM), and the reported 3-seed mean lacks error bars in tables.

Experimental setup

Datasets: DigitFive (MNIST, MNIST-M, SVHN, USPS, SynthDigits) Amazon Review (Books, DVDs, Electronics, Kitchen) public server dataset (unspecified)

Benchmarks: DigitFive Amazon Review

Metrics: T-Acc S-Acc G-Acc convergence rounds and time

Baselines: FOSDA SemiFDA AutoFedGP FedHEAL FedAVG FedProx FedProto

Builds on: FedAvg FedProx FedProto FOSDA FedHEAL AutoFedGP SemiFDA

Closest papers by meaning

• Enhancing Federated Domain Adaptation with Multi-Domain Prototype-Based Federated Fine-Tuning (ICLR 2025 · similarity 0.87)
• Principled Federated Domain Adaptation: Gradient Projection and Auto-Weighting (ICLR 2024 · similarity 0.86)
• UFDA: Universal Federated Domain Adaptation with Practical Assumptions (AAAI 2024 · similarity 0.86)
• Selective Aggregation for Low-Rank Adaptation in Federated Learning (ICLR 2025 · similarity 0.86)
• Knowledge Transfer via Compact Model in Federated Learning (Student Abstract) (AAAI 2024 · similarity 0.86)
• TTA-FedDG: Leveraging Test-Time Adaptation to Address Federated Domain Generalization (AAAI 2025 · similarity 0.86)