Workshop Day Schedule

Plenary: Privacy-aware Bayesian Networks

Cassio de Campos

Room: Guaratiba @ 2nd floor

Link to connect: https://uai2025.short.gy/SafeAI

Bayesian Networks are versatile probabilistic graphical models that enable meaningful knowledge representation and inferences. They have been proved effective across diverse domains, including healthcare, bioinformatics, economics, law, and image processing. As privacy concerns escalate, it becomes increasingly critical for publicly released models to safeguard sensitive information about the training data on which they were learned. Typically, released models do not prioritise privacy by design. The common current approach involves introducing noise into the model's parameters. While this idea can protect against tracing attacks, it also significantly impacts the model's utility. We present and discuss credal models as a practical solution for balancing privacy and utility. Credal models represent a set of standard precise models by having set-based parameter specifications. We discuss how credal models can disguise the original model, thereby reducing the probability of successful attacks, while achieving meaningful inferential results. Experiments illustrate the versatility of the idea and compare it against approaches based on noise.

Interactive Poster Session including Coffee served from 10:00-10:30

Location: foyer near the registration desk

• The Value of Recall in Extensive-Form Games (pdf)

Ratip Emin Berker (Carnegie Mellon University), Emanuel Tewolde (Carnegie Mellon University), Ioannis Anagnostides (Carnegie Mellon University), Tuomas Sandholm (Carnegie Mellon University), Vincent Conitzer (Carnegie Mellon University)

• HASARD: A Benchmark for Vision-Based Safe Reinforcement Learning in Embodied Agents (pdf) (poster)

Tristan Tomilin (Eindhoven University of Technology), Meng Fang (University of Liverpool), Mykola Pechenizkiy (Eindhoven University of Technology)

• DT-sampler: A SAT-based Decision Tree Ensemble (pdf)

Xiaotian Xue (The University of Tokyo), Chao Huang ( Rakuten Group, Inc. ), Koji Tsuda (The University of Tokyo), Diptesh Das (The University of Tokyo)

• SpaCE-VAE: Sparse and Confident Explanations using Variational Autoencoders (pdf) (poster)

Alexander Liu (TU Eindhoven), Sibylle Hess (TU Eindhoven)

• An Analysis of Robustness of Non-Lipschitz Networks (Extended Abstract) (pdf)

Maria-Florina Balcan (CMU), Avrim Blum (TTIC), DRAVYANSH SHARMA (TTIC), Hongyang Zhang (University of Waterloo)

Plenary: AI Security - Standards, STEADY and LASR

Paul Miller

Room: Guaratiba @ 2nd floor

Link to connect: https://uai2025.short.gy/SafeAI

In this talk I will first give a brief overview of the ETSI approach to developing standards for AI security through security-by-design. I will then discuss how we are applying an AI safety and assurance methodology, called ASLAM, to a specific use-case, namely unmanned ground vehicles. Finally, I will give a brief overview of a new UK initiative, the Laboratory of Artificial Intelligence Security Research and some of its activities.

Oral In-Person + Remote Presentations Session 15min + 5min Q&A

Chair: Mykola Pechenizkiy

Room: Guaratiba @ 2nd floor

Link to connect: https://uai2025.short.gy/SafeAI

• Decision Making under Imperfect Recall: Algorithms and Benchmarks (pdf)

Emanuel Tewolde (Carnegie Mellon University), Brian Zhang (Carnegie Mellon University), Ioannis Anagnostides ( Carnegie Mellon University ), Tuomas Sandholm (Carnegie Mellon University), Vincent Conitzer (Carnegie Mellon University)

• SafeFlowNet: Safe Control with Generative Flow Networks (pdf)

Yucheng Yang (Eindhoven University of Technology), Tianyi Zhou ( University of Maryland, College Park), Meng Fang (University of Liverpool / Eindhoven University of Technology), Mykola Pechenizkiy (Eindhoven University of Technology)

• Explaining Deep Learning Matching of Hand-Drawn Binary Symbols: A Visual Analysis with Grad-CAM on Cattle Brands (pdf) (slides)

Leandra Soares (Federal University of Goias), Marcos Medeiros (Federal University of Goias), Aldo Díaz-Salazar (Federal University of Goias), Edmundo Hoyle (GLOBO)

• Alignment metrics for the evaluation of observed autonomous systems (pdf) (slides)

Tommaso Mannucci (Autonomous Systems & Robotics, Unit DSS, TNO), Wouter Arink (Autonomous Systems & Robotics, Unit DSS, TNO), Johan van den Heuvel ( Autonomous Systems & Robotics, Unit DSS, TNO)

• Riemannian Manifold Learning for Stackelberg Games with Neural Flow Representations (pdf)

Larkin Liu (TU Munich), Kashif Rasul (Hugging Face, Inc.), Yutong Chao (TU Munich), Jalal Etesami (TU Munich)