Research map / Federated learning and privacy / Communication-efficient FL
Co-Dream: Collaborative Dream Synthesis over Decentralized Models
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 Co-Dream, a model-agnostic federated framework where clients collaboratively optimize random synthetic inputs ("dreams") using their local models and aggregate input gradients, instead of averaging model parameters.
Problem
Federated learning requires uniform model architectures across clients and communication proportional to model size, and knowledge-distillation alternatives are incompatible with secure aggregation.
Approach
Aggregate knowledge in data space: clients jointly optimize shared random inputs via DeepDream/DeepInversion-style entropy minimization plus BN-statistics and adversarial (JSD) regularizers, sending input gradients to the server; then all models are trained by distillation on the aggregated dreams and soft logits, with communication independent of model size.
Key results
- On CIFAR10 with non-IID alpha=10, Co-Dream reaches 62.2% average client accuracy vs 50.4% for AvgKD and 19.7% for independent training (heterogeneous models).
- On CIFAR10 with IID data, Co-Dream achieves 69.5% vs FedAvg 67.2% and centralized 65.7%, competitive with FedProx and SCAFFOLD.
- Membership inference (LiRA) on CIFAR10: CoDream-fast TPR@0.1% FPR is 39.36% vs 53.71% for FedAvg and 72.39% for single model, indicating lower leakage.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Radv to dream generation (ablation) | Co-Dream · CIFAR10, iid / α = 1 / α = 0.1 | accuracy: 65.7(0.2) -> 69.2(0.1); 58.4(1.3) -> 61.6(0.5); 42.0(1.4) -> 45.6(1.5) | Helped |
| Add Rbn to dream generation (ablation) | Co-Dream · CIFAR10, iid / α = 1 / α = 0.1 | accuracy: 51.2(6.1) -> 69.2(0.1); 33.1(7.1) -> 61.6(0.5); 24.1(5.2) -> 45.6(1.5) | Helped |
| Add collaborative optimization of dreams (ablation) | Co-Dream · CIFAR10, iid / α = 1 / α = 0.1 | accuracy: 64.4(0.5) -> 69.2(0.1); 58.4(1.4) -> 61.6(0.5); 30.8(3.2) -> 45.6(1.5) | Helped |
Surprising findings
- Co-Dream's communication cost is independent of model size, whereas FedAvg cost for VGG-11 is about 1013.6 MB per round, versus 600 MB for Co-Dream and 23.5 MB for CoDream-fast.
Limitations
- Does not beat state-of-the-art non-IID federated methods since it is not designed for non-IID challenges
- No formal privacy guarantee; privacy claims are empirical and via data-processing argument
- Client dropouts, stragglers and formal privacy guarantees left for future work
Experimental setup
Datasets: MNIST SVHN CIFAR10 CIFAR100
Benchmarks: MNIST SVHN CIFAR10 CIFAR100
Metrics: test accuracy balanced accuracy AUROC TPR at low FPR
Baselines: Centralized Independent FedAvg FedProx Moon AvgKD SCAFFOLD FedGen
Builds on: DeepDream DeepInversion FedAvg Fast-datafree knowledge distillation
Closest papers by meaning
- Overcoming Data and Model heterogeneities in Decentralized Federated Learning via Synthetic Anchors (ICML 2024 · similarity 0.86)
- FedQS: Optimizing Gradient and Model Aggregation for Semi-Asynchronous Federated Learning (NeurIPS 2025 · similarity 0.85)
- Reviving Stale Updates: Data-Free Knowledge Distillation for Asynchronous Federated Learning (NeurIPS 2026 · similarity 0.85)
- Decentralized Dynamic Cooperation of Personalized Models for Federated Continual Learning (NeurIPS 2025 · similarity 0.85)
- Decentralized Sporadic Federated Learning: A Unified Algorithmic Framework with Convergence Guarantees (ICLR 2025 · similarity 0.85)
- On the Necessity of Collaboration for Online Model Selection with Decentralized Data (NeurIPS 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.