Research map / Adversarial robustness and security / Certified and empirical robustness
A Little Is Enough: Circumventing Defenses For Distributed Learning
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 |
|---|---|---|---|
| Perturb parameters by 1.5 standard deviations across all workers | MNIST model with no defense · MNIST | accuracy: 96.1 -> 77.8 | Hurt |
| Perturb parameters by 1 standard deviation across all workers | CIFAR10 model with no defense · CIFAR10 | accuracy: 59.6 -> 20.9 | Hurt |
| Use the proposed small-perturbation attack against robust aggregation | CIFAR10 with 20% corrupted workers · CIFAR10 | accuracy: Krum: 66% drop; Bulyan: 52% drop; TrimmedMean: 45% drop | Hurt |
| Use the proposed small-perturbation attack against Krum | CIFAR10 distributed learning · CIFAR10 | accuracy: 33% drop | Hurt |
| Inject a sample backdoor using small perturbations | MNIST with Bulyan · MNIST test set, backdoor sample | accuracy: 96.1 -> 95.3 | Hurt |
| Inject a sample backdoor using small perturbations | CIFAR10 with Bulyan · CIFAR10 test set, backdoor sample | accuracy: 59.6 -> 54.2 | Hurt |
| Inject a pattern backdoor using small perturbations | MNIST with Bulyan · MNIST test set, pattern backdoor | backdoor target accuracy: 99.9 | Helped |
| Inject a pattern backdoor using small perturbations | CIFAR10 with Bulyan · CIFAR10 test set, pattern backdoor | backdoor target accuracy: 84.3 | Helped |
| Use parameter averaging instead of defense aggregation for the pattern-backdoor attack | MNIST model · MNIST test set, pattern backdoor | backdoor target accuracy: 36.9 | Hurt |
| Use parameter averaging instead of defense aggregation for the pattern-backdoor attack | CIFAR10 model · CIFAR10 test set, pattern backdoor | backdoor target accuracy: 7.3 | Hurt |
Closest papers by meaning
- On the Tension between Byzantine Robustness and No-Attack Accuracy in Distributed Learning (ICML 2025 · similarity 0.87)
- Exploit Gradient Skewness to Circumvent Byzantine Defenses for Federated Learning (AAAI 2025 · similarity 0.87)
- On the Robustness of Distributed Machine Learning Against Transfer Attacks (AAAI 2025 · similarity 0.87)
- High-Fidelity Gradient Inversion in Distributed Learning (AAAI 2024 · similarity 0.86)
- Byzantine-Robust Decentralized Learning via Remove-then-Clip Aggregation (AAAI 2024 · similarity 0.85)
- From Inexact Gradients to Byzantine Robustness: Acceleration and Optimization under Similarity (NeurIPS 2026 · 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.