Research map / Adversarial robustness and security / Backdoor and poisoning
Rethinking the Backdoor Attacks' Triggers: A Frequency Perspective
Yi Zeng, Won Park, Z. Morley Mao, Ruoxi Jia
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 |
|---|---|---|---|
| detect in DCT frequency domain instead of image domain | backdoor detector · CIFAR-10, BadNets | BDR (%): 1.38 -> 90.50 | Improved |
| detect in DCT frequency domain instead of image domain | backdoor detector · GTSRB, Troj-WM | BDR (%): 31.51 -> 100 | Improved |
| detect in DCT frequency domain instead of image domain | backdoor detector · CIFAR-10, l0 inv | BDR (%): 0.00 -> 100 | Improved |
| fine-tune CIFAR-10 detector with 200 GTSRB-distribution samples | DCT detector · GTSRB, transfer from CIFAR-10 | ACC (%): 68.23 -> 89.44 | Improved |
| fine-tune CIFAR-10 detector with 200 GTSRB-distribution samples | DCT detector · GTSRB, Blend trigger | BDR (%): 96.18 -> 45.67 | Hurt |
| fine-tune CIFAR-10 detector with 100 TSRD clean samples (size 200) | DCT detector · TSRD, BadNets | ACC (%): 61.17 -> 82.53 | Improved |
| train detector on combined CIFAR-10 and GTSRB instead of GTSRB alone | DCT detector · TSRD, Blend trigger | BDR (%): 88.94 -> 89.68 | Improved |
| add combined training data then tune on TSRD, versus no tuning | DCT detector · TSRD, Blend trigger | ACC (%): 64.18 -> 84.61 | Improved |
| use bilevel smooth trigger instead of naive low-pass filtered triggers | backdoor attack · CIFAR-10 | ASR (%): 75.54 -> 97.25 | Improved |
| use bilevel smooth trigger instead of naive low-pass filtered triggers | backdoor attack · CIFAR-10 | clean accuracy (%): 76.29 -> 84.54 | Improved |
| upgrade Meta Neural Analysis to consider smooth trigger generation | MNA defense · smooth-trigger poisoned classifiers | AUC: 0.0776 -> 0.694 | Improved |
| fine-tune detector with the smooth trigger instead of simple filtered patterns | DCT detector · unseen low-pass filtered triggers | detection rate (%): 89.37 | Improved |
Closest papers by meaning
- A Dual Stealthy Backdoor: From Both Spatial and Frequency Perspectives (AAAI 2024 · similarity 0.89)
- Rethinking Backdoor Attacks on Dataset Distillation: A Kernel Method Perspective (ICLR 2024 · similarity 0.89)
- Unveiling and Mitigating Backdoor Vulnerabilities based on Unlearning Weight Changes and Backdoor Activeness (NeurIPS 2024 · similarity 0.89)
- Inspecting Prediction Confidence for Detecting Black-Box Backdoor Attacks (AAAI 2024 · similarity 0.89)
- Demystifying Poisoning Backdoor Attacks from a Statistical Perspective (ICLR 2024 · similarity 0.88)
- LOTUS: Evasive and Resilient Backdoor Attacks through Sub-Partitioning (CVPR 2024 · similarity 0.88)
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.
