Research map / Adversarial robustness and security / Attacks on text and graphs
Analyzing Federated Learning through an Adversarial Lens
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Boost the malicious component of the update explicitly | Federated learning with one malicious agent · Fashion-MNIST | targeted misclassification confidence: 100% | Helped |
| Add local training loss and a penalty for distance from the previous benign update (ablation) | Targeted model poisoning · Fashion-MNIST | iterations selected by validation accuracy check: never -> 10 iterations out of 15 | Helped |
| Alternate targeted-objective optimization with stealth-objective optimization (ablation) | Stealthy model poisoning · Fashion-MNIST | stealth against update-distance checks | Helped |
| Use previous-step update estimation with pre-optimization correction (ablation) | Targeted model poisoning · Fashion-MNIST | targeted misclassification confidence at t=2, t=3, and t=4: 0.63 -> 0.82; 0.93 -> 0.98; 0.99 -> 1.0 | Helped |
| Use previous-step update estimation with pre-optimization correction (ablation) | Alternating minimization · Fashion-MNIST | targeted misclassification confidence at t=2, t=3, and t=4: 0.17 -> 0.47; 0.34 -> 0.89; 0.88 -> 1.0 | Helped |
| Use post-optimization correction instead of pre-optimization correction (ablation) | Model poisoning with update estimation · Federated learning | attack success and global-model convergence | Hurt |
| Use implicit boosting instead of explicit boosting (ablation) | Targeted model poisoning · Fashion-MNIST | iterations adversarial objective achieved: explicit boosting more successful; 4 of 10 iterations | Hurt |
| Use implicit boosting instead of explicit boosting (ablation) | Targeted model poisoning · Fashion-MNIST | optimization steps per time step: 2000 steps; 4× longer than a benign agent | Hurt |
| Add mislabeled copies of the target example rather than poison model updates (ablation) | Federated learning · Fashion-MNIST | targeted poisoning success: 1000 copies | Hurt |
| Use alternating minimization rather than targeted poisoning alone for ten target examples (ablation) | Model poisoning · Fashion-MNIST, r=10 | benign global-model accuracy while meeting the malicious objective | Helped |
| Use targeted model poisoning under Krum (ablation) | Krum aggregation · Fashion-MNIST | validation accuracy: 34 of 40 time steps | Hurt |
| Use alternating minimization under coordinate-wise median (ablation) | Coordinate-wise median aggregation · Fashion-MNIST | targeted attack effectiveness | Hurt |
Closest papers by meaning
- Model Poisoning Attacks to Federated Learning via Multi-Round Consistency (CVPR 2025 · similarity 0.92)
- FedREDefense: Defending against Model Poisoning Attacks for Federated Learning using Model Update Reconstruction Error (ICML 2024 · similarity 0.90)
- Tracing Back the Malicious Clients in Poisoning Attacks to Federated Learning (NeurIPS 2025 · similarity 0.89)
- Defending Against Sophisticated Poisoning Attacks with RL-based Aggregation in Federated Learning (AAAI 2025 · similarity 0.88)
- Dual Defense: Enhancing Privacy and Mitigating Poisoning Attacks in Federated Learning (NeurIPS 2024 · similarity 0.88)
- RFLPA: A Robust Federated Learning Framework against Poisoning Attacks with Secure Aggregation (NeurIPS 2024 · similarity 0.86)
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.