Research map / Adversarial robustness and security / Black-box and transfer attacks
Towards Understanding Adversarial Robustness of Optical Flow Networks
Simon Schrodi, Tonmoy Saikia, Thomas Brox
Check 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 |
|---|---|---|---|
| Use the deeper Robust FlowNetC encoder with a larger receptive field | FlowNetC · KITTI 2015 training, 102x102 handcrafted patch at worst location | EPE: 46.12 -> 13.08 | Improved |
| Replace attacked correlation features with unattacked features (ablation) | FlowNetC · KITTI 2015 training, 102x102 patch attack | EPE: 25.95 -> 12.67 | Improved |
| Replace attacked conv3<a,b> features with unattacked features (ablation) | FlowNetC · KITTI 2015 training, 102x102 patch attack | EPE: 25.95 -> 11.31 | Improved |
| Replace attacked conv_redir features with unattacked features (ablation) | FlowNetC · KITTI 2015 training, 102x102 patch attack | EPE: 25.95 -> 28.36 | Hurt |
| Replace the RAFT encoder with the original FlowNetC encoder and remove the context encoder (ablation) | RAFT · KITTI 2015 training, 102x102 handcrafted patch at worst location | EPE: 8.69 -> 19.12 | Hurt |
| Increase receptive field by increasing encoder depth (ablation) | FlowNetC variants · KITTI 2015 training, optimized patch attacks | EPE | Improved |
| Increase receptive field by increasing convolution dilation (ablation) | FlowNetC · KITTI 2015 training, optimized patch attacks | EPE: 18.81; 23.4; 22.66; 23.32; 22.3 | Mixed |
| Use the wrong input data normalization at inference (ablation) | FlowNetC · KITTI 2015 training, 102x102 optimized patch | EPE: 11.50 -> 14.52; 52.66 -> 94.51 | Hurt |
| Use a handcrafted high-frequency self-similar striped patch instead of an optimized patch | FlowNetC · KITTI 2015 training, 102x102 patch at worst location | EPE: 11.50 -> 46.12 | Hurt |
| Use untargeted I-FGSM perturbations | Robust FlowNetC, PWC-Net, and RAFT · KITTI 2015 training | EPE | Hurt |
| Use universal adversarial perturbations with small L-infinity norms | Optical flow networks · KITTI 2015 training | EPE | No clear effect |
| Fine-tune with adversarial data augmentation (ablation) | RAFT · KITTI 2015 training, unattacked images | EPE: 5.47 to 5.96, 7.13, 8.96 and 10.15 | Hurt |
Closest papers by meaning
- Attacking Optical Flow (ICCV 2019 · similarity 0.92)
- A Perturbation-Constrained Adversarial Attack for Evaluating the Robustness of Optical Flow (ECCV 2022 · similarity 0.90)
- When NAS Meets Robustness: In Search of Robust Architectures Against Adversarial Attacks (CVPR 2020 · similarity 0.86)
- Breaking Barriers in Physical-World Adversarial Examples: Improving Robustness and Transferability via Robust Feature (AAAI 2025 · similarity 0.84)
- SuperDeepFool: a new fast and accurate minimal adversarial attack (NeurIPS 2024 · similarity 0.83)
- Shape and Texture: What Influences Reliable Optical Flow Estimation? (CVPR 2025 · similarity 0.83)
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.
