Research map / Federated learning and privacy / Privacy attacks and secure learning
Poisoning with a Pill: Circumventing Detection in Federated Learning
Hanxi Guo, Hao Wang, Tao Song, Tianhang Zheng, Yang Hua, Haibing Guan, Xiangyu Zhang
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 poison pill augmentation to existing poisoning attacks | sign-flip, Trim, Krum, Min-Max attacks · Fashion-MNIST IID, 20% malicious, FedAvg and 5 defenses | error rate: avg +0.274, max +0.658 | Improved |
| add poison pill augmentation to existing poisoning attacks | sign-flip, Trim, Krum, Min-Max attacks · Fashion-MNIST non-IID, 20% malicious | error rate: avg +0.281, max +0.637 | Improved |
| add poison pill to sign-flipping attack against FedAvg | sign-flipping attack · Fashion-MNIST IID, FedAvg | error rate: 0.943 -> 0.667 | Hurt |
| add poison pill to Trim attack against FLTrust | Trim attack · Fashion-MNIST IID, FLTrust | error rate: 0.109 -> 0.576 | Improved |
| add poison pill to Krum attack against Multi-Krum | Krum attack · Fashion-MNIST IID, MKrum | error rate: 0.189 -> 0.715 | Improved |
| add poison pill to Krum attack against FLDetector | Krum attack · Fashion-MNIST IID, FLDetector | error rate: 0.786 -> 0.637 | Hurt |
| reduce malicious client proportion from 20% to 10% (ablation) | poison pill augmentation · Fashion-MNIST cross-silo | error rate increase: avg +0.144, max +0.403 (vs avg +0.274 at 20%) | Hurt |
| switch to cross-device FL (40% clients per round) (ablation) | poison pill augmentation · Fashion-MNIST cross-device | error rate increase: avg +0.279, max +0.639 | No clear effect |
| apply poison pill on CIFAR-10 with AlexNet | all four baseline attacks · CIFAR-10, 20% malicious, 30 clients | error rate: avg +0.288 | Improved |
| use approximate max pill search instead of approximate min pill search (ablation) | pill construction · pill search ablation (extended version) | attack error rate: better in 41 out of 56 cases | Improved |
| add extra local training rounds for malicious reference model (ablation) | pill poisoning · FLTrust server update, IID and non-IID | cosine similarity to FLTrust update | Improved |
| add pill adjustment (similarity and distance) (ablation) | poison pill injection · Multi-Krum distance score | distance score | Improved |
Closest papers by meaning
- Model Poisoning Attacks to Federated Learning via Multi-Round Consistency (CVPR 2025 · similarity 0.92)
- Tracing Back the Malicious Clients in Poisoning Attacks to Federated Learning (NeurIPS 2025 · similarity 0.92)
- FedREDefense: Defending against Model Poisoning Attacks for Federated Learning using Model Update Reconstruction Error (ICML 2024 · similarity 0.91)
- Dual Defense: Enhancing Privacy and Mitigating Poisoning Attacks in Federated Learning (NeurIPS 2024 · similarity 0.91)
- Defending Against Sophisticated Poisoning Attacks with RL-based Aggregation in Federated Learning (AAAI 2025 · similarity 0.88)
- RFLPA: A Robust Federated Learning Framework against Poisoning Attacks with Secure Aggregation (NeurIPS 2024 · similarity 0.88)
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.
