IPN Colloquium 19 10 2026 Marta Kwiatkowska

Artificial intelligence powered by neural networks has made significant advances in the past few years and its deployment is accelerating. However, neural network models are susceptible to adversarial perturbations of inputs, such as imperceptible manipulations of images or structural perturbations of graph data that can drastically change the network output. Beyond inference, data errors or poisoning attacks that inject malicious data into the training process can result in wrong or harmful model predictions, and in multi-agent reinforcement learning strategic manipulations can steer the learned policy towards undesirable goals. Undetected, such issues pose a significant risk to the safety and security of high-stakes applications. Formal methods, and particularly formal verification techniques, have become indispensable components of rigorous software engineering methodologies and are well placed to provide the needed assurance in the form of provable robustness guarantees on model predictions. Using illustrative examples, this lecture will overview recent progress and future challenges in research aimed at developing provably robust and strategyproof AI systems.