Physical AI Safety

Safe physical intelligence

What Makes Our Approach Distinctive

Physical AI systems operate through continuous interaction with the real world. They must understand human intent, perceive uncertain environments, predict future states, plan reliable actions, execute them through physical actuators, and continuously monitor whether their behavior remains safe.

Many Physical AI studies focus on robot architectures, demonstrations, or task-specific performance. We approach Physical AI safety from both modern learning theory and mathematical systems analysis.

Robust perception & State estimation

Reliable perception from noisy, incomplete, corrupted or adversarial sensor observations supporting Physical AI under sensor degradation, communication failure, unfamiliar environments and deliberate attack.

Selected Publications

- Kim et al., Local Geometry Attention for Time Series Forecasting under Realistic Corruptions, ICLR, 2026
- Kim et al., Towards Undetectable Adversarial Attack on Time Series Classification, Information Sciences, 2025
- Lee et al., Variational Cycle-Consistent Imputation Adversarial Networks for General Missing Patterns, Pattern Recognition, 2022
- Kim and Lee, Nonlinear Dynamic Projection for Noise Reduction of Dispersed Manifolds, IEEE TPAMI, 2014
- and more

Robust and reliable perception algorithms for Physical AI in uncertain and adversarial environments

Predictive world models & generative simulation

Learning temporal dynamics to predict future states and generate realistic trajectories supporting scenario synthesis, state reconstruction and simulation-based safety evaluation.

Selected Publications

- Park et al., TimeBridge: Better Diffusion Prior Design with Bridge Models for Time Series Generation, KDD, 2026
- Kim et al., Are Self-Attentions Effective for Time Series Forecasting?, NeurIPS, 2024
- Park et al., Fast Sharpness-Aware Training for Periodic Time Series Classification and Forecasting, Applied Soft Computing, 2023
- Kim et al., Self-Correcting Ensemble Using a Latent Consensus Model, Applied Soft Computing, 2016
- and more

Temporal dynamics modeling for state prediction, realistic trajectory generation, and simulation based safety evaluation

Safe planning & robust decision intelligence

Learning algorithms whose decisions stay dependable under adversarial perturbation, uncertainty and distribution shift — the learning theoretic foundation for reliable planning and risk-aware decisions.

Selected Publications

- Kim et al., Fantastic Robustness Measures: The Secrets of Robust Generalization, NeurIPS, 2023
- Lee et al., Implicit Jacobian Regularization Weighted with Impurity of Probability Output, ICML, 2023
- Lee et al., GradDiv: Adversarial Robustness of Randomized Neural Networks via Gradient Diversity Regularization, IEEE TPAMI, 2022
- Kim et al., Understanding Catastrophic Overfitting in Single-Step Adversarial Training, AAAI, 2021
- and more

Learning theoretic foundations for dependable planning and risk aware decisions

Stable control & verified optimization

Mathematical foundations for safe physical behavior: stability regions, unstable equilibria, nonlinear systems, trajectory-based optimization, homotopy methods and global optimization. This is what lets us state safe operating regions and failure boundaries mathematically.

Selected Publications

- Lee et al., Trajectory-Based Method for Multiple Solutions in Nonlinear Programming, IEEE TAC, 2004
- Lee, Optimization Framework for Computing the Controlling UEP, IEEE TAC, 2004
- Lee et al., Stability Regions of Non-Hyperbolic Dynamical Systems, IEEE TCAS-I, 2002
- Lee et al., Constructive Homotopy for Multiple DC Operating Points, IEEE TCAS-I, 2001
- Lee, Trajectory-Informed Search for Global Optimization, Journal of Global Optimization, 2007
- and more

Mathematical foundations for safe physical behavior via nonlinear optimization and stability analysis

runtime safety assurance & explainability

Monitoring deployed systems, identifying abnormal or unsafe behavior, explaining model decisions, and defending against security and privacy threats. The point is to connect detection and explanation to recovery, replanning and human oversight.

Selected Publications

- Jeong et al., JoCE: Joint Counterfactual Explanations for Interpretable Time Series Anomaly Detection, Pattern Recognition, 2026
- Choi et al.,Differentially Private Upsampling for Imbalanced Anomaly Detection , Engineering Applications of Artificial Intelligence, 2026
- Byun et al., Privacy-Preserving Inference Resistant to Model Extraction Attacks, Expert Systems with Applications, 2024
- and more

System-wide trustworthy AI foundations for Physical AI:
privacy, fairness, security, and representation stability

Trustworthy ai foundations

From Robust Learning to Stable Physical Intelligence

Trustworthy AI is a system-wide layer supporting every stage of Physical AI: learning from sensitive data, reducing algorithmic bias, protecting deployed systems, adapting to new environments, and maintaining stable representations.
Our research develops the learning algorithms and mathematical foundations required to make this closed-loop intelligence robust, stable, privacy-preserving, explainable, and verifiable. We focus particularly on reliable perception under realistic corruption, predictive and generative modeling, robust and certifiable decision intelligence, dynamical-system stability, global optimization, and runtime safety assurance.

Privacy-preserving & fair physical ai

Physical AI collects sensitive visual, acoustic, behavioral, spatial and industrial data. We develop privacy preserving learning and inference based on differential privacy and homomorphic encryption, alongside fair learning that reduces group level disparity.

Selected Publications

- Park et al., Multi-Class SVM with Differential Privacy, NeurIPS, 2025
- Choi et al., Safeguarding Retrieval Data against Membership Inference Attacks, EMNLP, 2025
- Park et al., Public Data Synthesis for Differentially Private Image Classification, CVPR, 2024
- Choi et al., Fair Sampling in Diffusion Models, AAAI, 2024
- Park et al., Differentially Private Sharpness-Aware Training, ICML, 2023
- and more

Privacy-preserving and fairness-aware learning for sensitive multimodal data generated by Physical AI systems

Stable & Robust Machine Learning

Mathematical and algorithmic methods that produce stable representations, robust cluster structures, transferable models, and reliable learning from limited or partially labeled data.

Selected Publications

- Kim et al., Voronoi Cell-Based Clustering Using a Kernel Support, IEEE TKDE, 2015
- Lee et al., Constructing Sparse Kernel Machines Using Attractors, IEEE TNN, 2009
- Kim and Lee, Clustering Based on Gaussian Processes, Neural Computation, 2007
- Lee and Lee, Equilibrium-Based SVM for Semi-Supervised Classification, IEEE TNN, 2007
- Lee et al., Dynamic Characterization of Cluster Structures for Robust SVC, IEEE TPAMI, 2006
- and more

Mathematical and algorithmic foundations for stable representations, robust clustering, model transferability, and reliable learning

data engineering

Earlier work in ensemble learning, large-scale data mining, Bayesian modeling, multimodal clustering, recommendation and information integration methodological foundations for heterogeneous-data analysis and data-driven decision support.

Selected Publications

- Park et al., Inductive Ensemble Clustering Using Kernel Support Matching, Electronics Letters, 2017
- Kim et al., Improved Churn Prediction by Analyzing a Large Network, Expert Systems with Applications, 2014
- Park et al., Transductive Bayesian Regression via Manifold Learning, Expert Systems with Applications, 2012
- Heo et al., Credit Delinquent Prediction Using Locally Transductive MLP, Neurocomputing, 2009
- Cho et al., A Tandem Clustering Process for Multimodal Datasets, European Journal of Operational Research, 2006
- and more

Foundations for heterogeneous data analysis and data driven decision support

Page updated

Google Sites

Report abuse