Research map / Adversarial robustness and security / Black-box and transfer attacks
Ripple Perturbations Through Structure: Likelihood-Constrained Adversarial Attacks on Heterogeneous Tabular Data
Zhengjie Zhou, Jiahuan Yan, Boqun Ma, Weiwei Feng, Tengfei LIU, Weiqiang Wang
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 |
|---|---|---|---|
| add structure-aware ripple perturbation with SCM likelihood constraint (LCSA) over structure-blind attacks | tabular attacks · URL TabTransformer, L2=0.5 & gamma=0.5 | robust accuracy: 11.3 (CAPGD) -> 2.1 (LCSA) | Improved |
| add LCSA structural budget on top of geometric budget | attack · Synthetic Chain, L2=0.5 | robust accuracy: 8.3 (CAPGD) -> 13.9 (LCSA) | Hurt |
| add LCSA structural constraint | attack · Synthetic Chain, L2=0.5 | SCE: 1.59 (CAPGD) -> 0.14 (LCSA) | Improved |
| use a forced DAG for the Cyclic topology | LCSA · Synthetic Cyclic | SCE: 2.15 (CAPGD) -> 0.64 (LCSA) | Improved |
| increase mechanism noise in the learned SCM (ablation) | LCSA · Confounder topology | robust accuracy: 22.2% at sigma = 0.1 | No clear effect |
| drop 40% of edges in the learned structure (ablation) | LCSA · Confounder topology | robust accuracy: 29.5% at p = 0.4 | No clear effect |
| remove edges from learned SCM on Chain topology (ablation) | LCSA · Chain topology | robust accuracy | Hurt |
| reverse edge directions in learned SCM (ablation) | LCSA · Confounder topology | robust accuracy: stays below 35% as SCE rises to 9.2 | No clear effect |
| tighten structural and L2 budgets to 0.1 (ablation) | LCSA · URL | robust accuracy: 4.5 (L2=0.3, gamma=0.3) vs 25.4 (L2=0.1, gamma=0.1) | Hurt |
| adversarially train models on LCSA examples | TabTransformer · URL | robust accuracy against LCSA: 2.1 -> 88.9 | Improved |
| transfer LCSA examples to non-differentiable XGBoost | LCSA · URL TabNet to XGBoost | robust accuracy: 37.1 (MOEVA) -> 32.4 (LCSA) | Improved |
| use LCSA's structure-consistent perturbations against an Isolation Forest detector | LCSA · URL | anomaly detection ratio: 32.7 (CaFA) -> 19.4 (LCSA) | Improved |
Closest papers by meaning
- Constrained Adaptive Attack: Effective Adversarial Attack Against Deep Neural Networks for Tabular Data (NeurIPS 2024 · similarity 0.87)
- TabularBench: Benchmarking Adversarial Robustness for Tabular Deep Learning in Real-world Use-cases (NeurIPS 2024 · similarity 0.87)
- RAD-TFM: Robust and Domain-Adapted Tabular Foundation Models (NeurIPS 2026 · similarity 0.87)
- Contrasting Adversarial Perturbations: The Space of Harmless Perturbations (AAAI 2025 · similarity 0.85)
- Tabular Foundation Model for Generative Modelling (NeurIPS 2026 · similarity 0.84)
- Causal Adversarial Perturbations for Individual Fairness and Robustness in Heterogeneous Data Spaces (AAAI 2024 · similarity 0.84)
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.
