Research map / Adversarial robustness and security / Certified and empirical robustness
Bridging Mode Connectivity in Loss Landscapes and Adversarial Robustness
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 |
|---|---|---|---|
| Connect two backdoored models using limited bonafide data | Backdoored VGG on CIFAR-10 · Single-target backdoor, 50 bonafide samples | clean accuracy: 63% vs 46% | Helped |
| Connect two backdoored models using limited bonafide data | Backdoored ResNet on SVHN · Single-target backdoor, 50 bonafide samples | backdoor accuracy: 16% vs 60% | Helped |
| Prune weights and fine-tune with bonafide data | Backdoored VGG on CIFAR-10 · Single-target backdoor, 50 bonafide samples | backdoor accuracy: 82% vs 2.5% | Hurt |
| Connect two error-injected models using limited bonafide data | Error-injected VGG on CIFAR-10 · Error injection, 50 bonafide samples | clean accuracy: 88% vs 82% | Helped |
| Connect two error-injected models using limited bonafide data | Error-injected VGG on CIFAR-10 · Error injection, 50 bonafide samples | injection accuracy: 0% vs 25% | Helped |
| Reduce path-training bonafide data from 2500 to 50 samples (ablation) | Path connection for backdoor repair · CIFAR-10 (VGG), single-target backdoor | clean accuracy: 88% -> 63% | Hurt |
| Reduce path-training bonafide data from 2500 to 50 samples (ablation) | Path connection for backdoor repair · CIFAR-10 (VGG), single-target backdoor | backdoor accuracy: 1.1% -> 2.5% | Mixed |
| Use standard-loss path connection between regular and adversarially trained models | Regular-to-robust model path · Evasion attacks | robustness loss | Hurt |
| Use adversarial training rather than standard-loss path training | Path connection between model pairs · Evasion attacks | robustness loss | Helped |
| Ensemble models from regular or robust connection paths | Path-connected models · Evasion attacks | robustness against evasion attacks | No clear effect |
Closest papers by meaning
- Input Space Mode Connectivity in Deep Neural Networks (ICLR 2025 · similarity 0.89)
- Optimizing Mode Connectivity via Neuron Alignment (NeurIPS 2020 · similarity 0.87)
- Revisiting Residual Networks for Adversarial Robustness (CVPR 2023 · similarity 0.87)
- Exploring Architectural Ingredients of Adversarially Robust Deep Neural Networks (NeurIPS 2021 · similarity 0.87)
- Expressive Losses for Verified Robustness via Convex Combinations (ICLR 2024 · similarity 0.85)
- On Linear Mode Connectivity of Mixture-of-Experts Architectures (NeurIPS 2025 · 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.