Research map / Adversarial robustness and security / Certified and empirical robustness
Robust Fine-Tuning from Non-Robust Pretrained Models: Mitigating Suboptimal Transfer With Epsilon-Scheduling
Jonas Ngnawe, Maxime Heuillet, Sabyasachi Sahoo, Yann Pequignot, Ola Ahmad, Audrey Durand, Frederic Precioso, Christian Gagné
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 epsilon-scheduling (zero-epsilon warmup then linear ramp) to robust fine-tuning | RFT-fix, SWIN · Aircraft (eps=4/255) | clean accuracy: 7.70 -> 73.80 | Improved |
| add epsilon-scheduling to robust fine-tuning | RFT-fix, ClipCNX · Cars (eps=8/255) | clean accuracy: 1.30 -> 90.00 | Improved |
| add epsilon-scheduling | RFT-fix, six non-robust backbones · 5 datasets avg (eps=8/255) | expected adversarial accuracy: 10.12 -> 34.38 (ViT) | Improved |
| add epsilon-scheduling | RFT-fix, six backbones · 30 configurations (eps=4/255) | adversarial accuracy: fix better adversarial accuracy in 9 of 30 | Mixed |
| add epsilon-scheduling | RFT-fix, six backbones · 30 configurations (eps=8/255) | adversarial accuracy: scheduler better in 28 of 30 | Improved |
| add epsilon-scheduling to robust pretrained backbone | RFT-fix, RobSWIN · Cars (eps=8/255) | expected adversarial accuracy: 50.62 -> 63.88 | Improved |
| add epsilon-scheduling to robust pretrained backbone | RFT-fix, RobCNX · Aircraft (eps=4/255) | adversarial accuracy: 62.00 -> 49.70 | Hurt |
| add zero-epsilon warmup phase T1=12 (vs T1=0) with T2=30 (ablation) | Linear warmup, SWIN · Cars (eps=4/255) | expected adversarial accuracy: 63.41 -> 67.09 | Improved |
| use gradual ramp (T2=37) instead of abrupt switch (T1=T2=12) (ablation) | Delayed switch, SWIN · Cars (eps=4/255) | clean accuracy: 1.90 -> 84.70 | Improved |
| lengthen ramp from T2=37 to T2=50 at T1=12 (ablation) | Epsilon-Scheduling, SWIN · Cars (eps=4/255) | adversarial accuracy: 43.20 -> 35.80 | Hurt |
| use scheduled epsilon instead of uniform random epsilon (ablation) | Uniform random epsilon, SWIN · Cars (eps=8/255) | expected adversarial accuracy: 30.16 -> 53.57 | Improved |
| use scheduled epsilon instead of uniform random epsilon (ablation) | Uniform random epsilon, ClipCNX · Aircraft (eps=4/255) | expected adversarial accuracy: 5.78 -> 67.88 | Improved |
Closest papers by meaning
- On the Robustness Tradeoff in Fine-Tuning (ICCV 2025 · similarity 0.92)
- Adversarial Robustness: From Self-Supervised Pre-Training to Fine-Tuning (CVPR 2020 · similarity 0.90)
- Initialization Matters for Adversarial Transfer Learning (CVPR 2024 · similarity 0.88)
- Understanding Overadaptation in Supervised Fine-Tuning: The Role of Ensemble Methods (ICML 2025 · similarity 0.87)
- PACE: Marrying generalization in PArameter-efficient fine-tuning with Consistency rEgularization (NeurIPS 2024 · similarity 0.87)
- Mechanistically analyzing the effects of fine-tuning on procedurally defined tasks (ICLR 2024 · similarity 0.87)
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.
