Research map / Adversarial robustness and security / Attacks on text and graphs
Houdini: Fooling Deep Structured Visual and Speech Recognition Models with Adversarial Examples
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 |
|---|---|---|---|
| use Houdini loss instead of MSE for untargeted attack (ablation) | pose estimation attack (Hourglass) · MPII val | PCKh0.5 after 100 iterations: 24.12 -> 0.57 | Helped |
| use Houdini instead of MSE at convergence for untargeted pose attack (ablation) | pose estimation attack · MPII val | perceptibility: up to 50% less perceptible | Helped |
| use Houdini instead of NLL, untargeted segmentation attack at 50% mIoU drop (ablation) | Dilation10 attack · Cityscapes val | perceptibility: 0.0037 -> 0.0026 | Helped |
| use Houdini instead of NLL, untargeted segmentation attack at convergence (ablation) | Dilation10 attack · Cityscapes val | perceptibility: 0.0117 -> 0.0095 | Helped |
| use Houdini instead of NLL, targetted segmentation attack at 50% mIoU drop (ablation) | Dilation10 attack · Cityscapes val | perceptibility: 0.0074 -> 0.0054 | Helped |
| use Houdini instead of NLL, targetted segmentation attack at convergence (ablation) | Dilation10 attack · Cityscapes val | perceptibility: 0.0389 -> 0.0392 | No clear effect |
| use Houdini instead of NLL, untargeted segmentation (ablation) | Dilation10 attack · Cityscapes val | SSIM @mIoU/2: 0.9989 -> 0.9995 | Helped |
| use Houdini instead of NLL, targetted segmentation (ablation) | Dilation10 attack · Cityscapes val | SSIM @mIoU/2: 0.9972 -> 0.9975 | Helped |
| use Houdini instead of CTC loss to craft adversarial audio, eps=0.05 (ablation) | DeepSpeech-2 style ASR · Librispeech clean test | WER: 20 -> 46.5 | Helped |
| use Houdini instead of CTC loss, eps=0.05 (ablation) | DeepSpeech-2 style ASR · Librispeech clean test | CER: 2.5 -> 4.5 | Helped |
| use Houdini instead of CTC loss, eps=0.3 (ablation) | DeepSpeech-2 style ASR · Librispeech clean test | WER: 68 -> 96.1 | Helped |
| use Houdini instead of CTC loss, eps=0.1 (ablation) | DeepSpeech-2 style ASR · Librispeech clean test | WER: 29.8 -> 66.5 | Helped |
Closest papers by meaning
- Imperceptible, Robust, and Targeted Adversarial Examples for Automatic Speech Recognition (ICML 2019 · similarity 0.86)
- Weighted-Sampling Audio Adversarial Example Attack (AAAI 2020 · similarity 0.86)
- GAT: Generative Adversarial Training for Adversarial Example Detection and Robust Classification (ICLR 2020 · similarity 0.86)
- Adversarial Examples Improve Image Recognition (CVPR 2020 · similarity 0.85)
- High Fidelity Speech Synthesis with Adversarial Networks (ICLR 2020 · similarity 0.85)
- Adversarial Examples Can Be Effective Data Augmentation for Unsupervised Machine Learning (AAAI 2022 · 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.