Research map / Federated learning and privacy / Privacy attacks and secure learning
FedIPR: Ownership Verification for Federated Deep Neural Network Models
Bowen Li, Lixin Fan, Hanlin Gu, Jie Li, Qiang Yang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add feature-based watermark embedding in FedAvg training (N=100 vs baseline) | AlexNet on CIFAR10, 20 clients · CIFAR10 main accuracy | accuracy: 91.72% -> 90.62% | Hurt |
| add feature-based watermark embedding (N=100) vs FedAvg | ResNet18 on CIFAR100, 20 clients · CIFAR100 main accuracy | accuracy: 76.52% -> 74.29% | Hurt |
| add backdoor-based watermarks (N_T=50) vs FedAvg | AlexNet on CIFAR10, 20 clients · CIFAR10 main accuracy | accuracy: 91.72% -> 91.69% | No clear effect |
| embed watermarks in normalization layer scale parameters instead of convolution weights (ablation) | FedIPR feature-based watermarks · fine-tuning attack, AlexNet CIFAR10 | feature detection rate eta_F: remains at 100% vs drops significantly for convolution layer | Improved |
| embed watermarks in normalization layer rather than convolution layer (ablation) | FedIPR feature-based watermarks · pruning attack, AlexNet CIFAR10 | feature detection rate eta_F | Improved |
| increase total watermark bits KN beyond channel count M (ablation) | FedIPR feature-based watermarks · feature detection rate vs bit length | feature detection rate eta_F: 100% -> about 80% | Hurt |
| increase feature watermark length N per client from 50 to 500 (beta=0.1, non-iid) (ablation) | ResNet on CIFAR100 · non-iid Dirichlet beta=0.1 | feature detection rate eta_F: 100.0% -> 78.45% | Hurt |
| increase feature watermark length N per client from 50 to 300 (beta=0.1, non-iid) (ablation) | AlexNet on CIFAR10 · non-iid Dirichlet beta=0.1 | feature detection rate eta_F: 100% -> 76.97% | Hurt |
| increase backdoor trigger count per client from 50 to 300 (ablation) | FedIPR backdoor watermarks, AlexNet CIFAR10 20 clients · CIFAR10 backdoor detection | backdoor detection rate eta_T: 99.34% -> 98.85% | No clear effect |
| replace FedAvg with Trim-mean defensive aggregation (ablation) | FedIPR backdoor watermarks · CIFAR10 defensive aggregation | backdoor detection rate eta_T: 98.82% -> 63.25% | Hurt |
| replace FedAvg with Bulyan defensive aggregation (ablation) | FedIPR feature watermarks · CIFAR10 defensive aggregation | feature detection rate eta_F: 100% -> 98.75% | No clear effect |
| lower client sampling ratio c to 0.25 (ablation) | FedIPR · client selection, 10 clients | detection rates eta_F and eta_T | No clear effect |
Closest papers by meaning
- ZeroMark: Towards Dataset Ownership Verification without Disclosing Watermark (NeurIPS 2024 · similarity 0.88)
- MFL-Owner: Ownership Protection for Multi-modal Federated Learning via Orthogonal Transform Watermark (AAAI 2025 · similarity 0.87)
- FedGMark: Certifiably Robust Watermarking for Federated Graph Learning (NeurIPS 2024 · similarity 0.86)
- EntropyMark: Towards More Harmless Backdoor Watermark via Entropy-based Constraint for Open-source Dataset Copyright Protection (CVPR 2025 · similarity 0.85)
- Revisiting the Information Capacity of Neural Network Watermarks: Upper Bound Estimation and Beyond (AAAI 2024 · similarity 0.85)
- Safe and Robust Watermark Injection with a Single OoD Image (ICLR 2024 · similarity 0.83)
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.
