Research map / Federated learning and privacy / Privacy attacks and secure learning
SABRE-FL: Selective and Accurate Backdoor Rejection for Federated Prompt Learning
Momin Ahmad Khan, Yasra Chandio, Fatima M. Anwar
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 SABRE-FL detector filtering versus no defense | federated prompt learning with CLIP · FGVC Aircraft, 25% malicious | backdoor accuracy: 93.9 -> 7.6 | Improved |
| SABRE-FL versus FLAME, the strongest baseline | SABRE-FL · FGVC Aircraft | backdoor accuracy: 16.4 -> 7.6 | Improved |
| use SABRE-FL, effect on clean accuracy | SABRE-FL · five datasets | clean accuracy: e.g. Flowers 77.9 -> 76.6, Pets 94.2 -> 94.5 | No clear effect |
| use Norm Bounding defense | FL prompt learning · Pets | backdoor accuracy: 16.3 -> 22.5 | Hurt |
| scale to 32 clients with SABRE-FL (ablation) | SABRE-FL · FGVC Aircraft | backdoor accuracy: 89.9 -> 24.7 | Improved |
| leave more malicious clients unfiltered (1 to 4) (ablation) | SABRE-FL · DTD | backdoor accuracy: 11.3 -> 25.5 | Hurt |
| increase number of shots to 16 (no defense) (ablation) | attack · FGVC Aircraft and Food101 | backdoor accuracy: over 85% at 16 shots | Hurt |
| add SABRE-FL across 2 to 16 shots (ablation) | SABRE-FL · five datasets | backdoor accuracy: under 5% | Improved |
| strong non-IID data (Dirichlet alpha 0.1) (ablation) | SABRE-FL · FGVC Aircraft | backdoor accuracy: alpha 0.9: 86.7 -> 0.1; alpha 0.1: 81.3 -> 89.6 | Mixed |
| increase malicious client proportion to 50% or more (ablation) | attack without defense · most datasets | backdoor accuracy: exceeds 80% | Hurt |
| increase trigger optimization epochs from 1 to 5 (no defense) (ablation) | attack · Caltech | backdoor accuracy: 31.6 -> 58.4 | Hurt |
Closest papers by meaning
- ShadowFPT: Backdooring Federated Prompt Tuning via Shadow Triggers (NeurIPS 2026 · similarity 0.91)
- Leveraging Soft Prompts for Privacy Attacks in Federated Prompt Tuning (NeurIPS 2026 · similarity 0.89)
- FLiP: Towards Comprehensive and Reliable Evaluation of Federated Prompt Learning (NeurIPS 2025 · similarity 0.88)
- Label-Free Backdoor Attacks in Vertical Federated Learning (AAAI 2025 · similarity 0.87)
- MARS: A Malignity-Aware Backdoor Defense in Federated Learning (NeurIPS 2025 · similarity 0.86)
- Resisting Backdoor Attacks in Federated Learning via Bidirectional Elections and Individual Perspective (AAAI 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.
