Research map / Adversarial robustness and security
Certified and empirical robustness: research map
1,356 papers on Certified and empirical robustness in Adversarial robustness and security, from ICML, NeurIPS, ICLR, CVPR, AAAI, ICCV and ECCV, and the journals TPAMI, IJCV and JMLR (2016–2026), grouped into 4 clusters and 16 approaches. The busiest year so far is 2022.
What it is
Tiny, carefully chosen changes to an input (adversarial examples) can fool a network. This topic trains models to resist them (adversarial training), evaluates defences honestly, and proves guarantees that no attack within a budget can succeed (certified robustness).
Terms you will see on this page:
- Adversarial example
- an input changed imperceptibly to cause a wrong prediction.
- Adversarial training
- training on adversarial examples so the model resists them.
- Randomized smoothing
- certifying robustness by averaging predictions over random noise.
Start reading with the most cited papers here:
- DeepFool: A Simple and Accurate Method to Fool Deep Neural Networks (CVPR 2016 · 5,359 citations)
- mixup: Beyond Empirical Risk Minimization (ICLR 2018 · 4,723 citations)
Within Adversarial robustness and security, its share shrank from 38.9% in 2023–24 to 23.5% in 2025–26 (303 → 231 papers at ICML, NeurIPS, ICLR, CVPR and AAAI, the venues with data for all four years).
Explore Certified and empirical robustness in the interactive map
Working on something in this topic? Describe your idea in scime atlas to see which approach it falls under, the closest papers by meaning and how crowded the spot has become.
What changed recently
Approaches whose share of the topic moved most between 2023–24 and 2025–26 (ICML, NeurIPS, ICLR, CVPR and AAAI):
- Adversarial training with feature consistency rising: 14% → 20% of the topic (43 → 46 papers)
- Robust overfitting and regularization rising: 5% → 10% of the topic (16 → 22 papers)
- Theory of adversarial risk rising: 5% → 8% of the topic (14 → 18 papers)
- Adversarial training refinements falling: 11% → 7% of the topic (32 → 16 papers)
- Neural network robustness verification falling: 9% → 6% of the topic (27 → 13 papers)
Papers from 2025–2026 that are already cited most:
- AttackBench: Evaluating Gradient-based Attacks for Adversarial Examples (AAAI 2025 · 18 citations)
- 3SAT: A Simple Self-Supervised Adversarial Training Framework (AAAI 2025 · 3 citations)
- Can Go AIs Be Adversarially Robust? (AAAI 2025 · 3 citations)
- MOS-Attack: A Scalable Multi-objective Adversarial Attack Framework (CVPR 2025 · 3 citations)
Approaches and key papers
Robustness trade-offs and evaluation · 609 papers
Keywords: adversarially, generalization, overfitting
Approaches in this cluster:
- Robustness-accuracy trade-off analysis (181 papers)
Study adversarial training and its evaluation, including trade-offs with accuracy and weak-attack pitfalls. - Attack evaluation and defense critique (137 papers)
Analyze adversarial examples, attack strength and why defenses fail. - Adversarial training refinements (124 papers)
Reweight examples and classes or tune training to improve robustness and convergence. - Robust overfitting and regularization (101 papers)
Explain and reduce robust overfitting and improve generalization of adversarially trained models. - Theory of adversarial risk (66 papers)
Formalize adversarially robust learning, Bayes classifiers and PAC-Bayes bounds.
Most cited and most cited since 2024:
- Towards Deep Learning Models Resistant to Adversarial Attacks (ICLR 2018 · 1,508 citations)
- Theoretically Principled Trade-off between Robustness and Accuracy (ICML 2019 · 1,119 citations)
- Towards Understanding and Improving Adversarial Robustness of Vision Transformers (CVPR 2024 · 19 citations)
- Revisiting Adversarial Training Under Long-Tailed Distributions (CVPR 2024 · 18 citations)
Verification and provable robustness · 319 papers
Keywords: verification, bounds, worst case
Approaches in this cluster:
- Provable robust learning guarantees (125 papers)
Derive worst-case guarantees and relaxations for adversarially robust training. - Neural network robustness verification (91 papers)
Certify robustness with semidefinite, probabilistic or bound-propagation verifiers. - Decision-boundary geometry of robustness (70 papers)
Analyze decision boundaries and universal perturbations to explain adversarial vulnerability. - Robustness to poisoning and multiple attacks (33 papers)
Benchmark and certify defenses against poisoning and multiple threat models.
Most cited and most cited since 2024:
- DeepFool: A Simple and Accurate Method to Fool Deep Neural Networks (CVPR 2016 · 5,359 citations)
- Provable Defenses against Adversarial Examples via the Convex Outer Adversarial Polytope (ICML 2018 · 674 citations)
- Promoting Counterfactual Robustness through Diversity (AAAI 2024 · 8 citations)
- Clarifying the Behavior and the Difficulty of Adversarial Training (AAAI 2024 · 4 citations)
Adversarial training · 315 papers
Keywords: cifar, 10, imagenet
Approaches in this cluster:
- Adversarial training recipes (80 papers)
Improve robust training through fast training, pretraining, teacher guidance and local features. - Regularized defenses against attacks (45 papers)
Add regularizers, detection or domain adaptation to harden networks against adversarial examples. - Augmentation and diffusion purification (37 papers)
Use mixup, augmentation and diffusion models to purify or robustify inputs. - Adversarial training with feature consistency (153 papers)
Improve robustness via information-theoretic, distillation and consistency regularizers.
Most cited and most cited since 2024:
- mixup: Beyond Empirical Risk Minimization (ICLR 2018 · 4,723 citations)
- Fast is better than free: Revisiting adversarial training (ICLR 2020 · 576 citations)
- Revisiting Adversarial Training at Scale (CVPR 2024 · 27 citations)
- AttackBench: Evaluating Gradient-based Attacks for Adversarial Examples (AAAI 2025 · 18 citations)
Certified robustness · 113 papers
Keywords: smoothing, randomized, certified robustness
Approaches in this cluster:
- Randomized smoothing certification (53 papers)
Certify robustness by smoothing classifiers with noise, extending to new norms and tasks. - Certified training and ensembles (36 papers)
Train models to be certifiably robust, including ensembles and bounded-support settings. - Provable defenses and their limits (24 papers)
Scale, break or extend certified defenses using generated data and partition aggregation.
Most cited and most cited since 2024:
- Certified Adversarial Robustness via Randomized Smoothing (ICML 2019 · 587 citations)
- Provably Robust Deep Learning via Adversarially Trained Smoothed Classifiers (NeurIPS 2019 · 261 citations)
- Diffusion Models are Certifiably Robust Classifiers (NeurIPS 2024 · 14 citations)
- MMCert: Provable Defense against Adversarial Attacks to Multi-modal Models (CVPR 2024 · 7 citations)
Related topics in Adversarial robustness and security
- Backdoor and poisoning (385)
- Black-box and transfer attacks (513)
- Attacks on text and graphs (1,300)
