Research map / Adversarial robustness and security / Backdoor and poisoning
RIBAC: Towards Robust and Imperceptible Backdoor Attack against Compact DNN
Huy Phan, Cong Shi, Yi Xie, Tianfang Zhang, Zhuohang Li, Tianming Zhao, Jian Liu, Yan Wang, Yingying Chen, Bo Yuan
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 |
|---|---|---|---|
| learn pruning and backdoor jointly in parallel (P+B) instead of backdoor injection then pruning (B->P) (ablation) | RIBAC · CIFAR-10 (16x, all-to-all) | attack success rate (%): 69.50 -> 100.00 | Improved |
| learn pruning and backdoor jointly in parallel (P+B) instead of B->P (ablation) | RIBAC · Tiny ImageNet (32x, all-to-all) | attack success rate (%): 0.50 -> 99.28 | Improved |
| use parallel P+B instead of pruning a clean pre-trained network then WaNet backdoor training (ablation) | RIBAC · Tiny ImageNet (32x, all-to-all) | attack success rate (%): 53.06 -> 99.28 | Improved |
| use parallel P+B instead of random-init sparse model with WaNet training (ablation) | RIBAC · CIFAR-10 (32x, all-to-all) | clean accuracy (%): 89.25 -> 92.39 | Improved |
| use parallel P+B instead of B->P (ablation) | RIBAC · CelebA (16x, all-to-all) | clean accuracy (%): 27.16 -> 81.68 | Improved |
| use RIBAC (backdoor plus pruning) instead of L1 pruning alone | L1 pruning · CIFAR-10 (32x) | clean accuracy (%): 91.51 -> 91.77 | No clear effect |
| use RIBAC instead of L1 pruning alone | L1 pruning · Tiny ImageNet (32x) | clean accuracy (%): 55.64 -> 53.98 | No clear effect |
| use RIBAC instead of importance-score pruning alone | importance-score pruning · CIFAR-10 (32x) | clean accuracy (%): 93.05 -> 91.77 | Hurt |
| use RIBAC instead of existing backdoor methods without compression | WaNet/LIRA/WB · Tiny ImageNet | clean accuracy and attack success rate: up to 2.76% clean accuracy and 40.64% attack success rate increase | Improved |
| use RIBAC instead of WaNet in all-to-all mode | WaNet · Tiny ImageNet (all-to-all) | attack success rate (%): 58.00 -> 99.31 | Improved |
| increase compression ratio from 2x to 32x | RIBAC (all-to-one) · CIFAR-10 | clean accuracy (%): 94.35 -> 91.77 | Hurt |
| increase compression ratio from 2x to 32x | RIBAC (all-to-all) · GTSRB | attack success rate (%): 99.98 -> 98.34 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- WaNet — WaNet - Imperceptible Warping-based Backdoor Attack (ICLR 2021) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- From Toxic to Trustworthy: Using Self-Distillation and Semi-supervised Methods to Refine Neural Networks (AAAI 2024 · similarity 0.86)
- Uncovering, Explaining, and Mitigating the Superficial Safety of Backdoor Defense (NeurIPS 2024 · similarity 0.85)
- Inspecting Prediction Confidence for Detecting Black-Box Backdoor Attacks (AAAI 2024 · similarity 0.85)
- REFINE: Inversion-Free Backdoor Defense via Model Reprogramming (ICLR 2025 · similarity 0.85)
- Conditional Backdoor Attack via JPEG Compression (AAAI 2024 · similarity 0.85)
- Efficient Backdoor Attacks for Deep Neural Networks in Real-world Scenarios (ICLR 2024 · 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.
