MiniFool -- Physics-Constraint-Aware Minimizer-Based Adversarial Attacks in Deep Neural Networks

A team of researchers has introduced MiniFool, a physics-inspired algorithm designed to probe the vulnerabilities of deep neural networks used in particle and astroparticle physics classification tasks [1]. The algorithm, detailed in a paper revised on 16 June 2026, was initially developed for the search for astrophysical tau neutrinos at the IceCube Neutrino Observatory [2]. It has since been applied to the MNIST data set and Open Data from the CMS experiment at the Large Hadron Collider to demonstrate its general applicability [2]. The work is led by Christopher Wiebusch and collaborators [1]. MiniFool operates by minimizing a cost function that combines a χ²-based test statistic with the deviation from a desired target score [2]. The test statistic quantifies the probability of the perturbations applied to the data based on experimental uncertainties [2]. This approach differs from conventional adversarial attacks by incorporating physical constraints directly into the perturbation process. The study finds that the likelihood of a flipped classification differs for events that were initially correctly and incorrectly classified [2]. By testing changes in classifications as a function of an attack parameter that scales the experimental uncertainties, the robustness of the network's decision can be quantified [2]. This method also allows for testing the robustness of classifications on unlabeled experimental data [2]. Neutrino physics provides a demanding testbed for such algorithms. Neutrino oscillation, a quantum mechanical phenomenon where a neutrino created with a specific lepton flavor can later be measured as a different flavor, was experimentally confirmed by the Super-Kamiokande and Sudbury Neutrino Observatories, a discovery recognized with the 2015 Nobel Prize in Physics [5]. The IceCube Observatory, for which MiniFool was first developed, searches for high-energy astrophysical neutrinos, including the tau flavor, which are critical for understanding cosmic accelerators [2]. The robustness testing enabled by MiniFool arrives as deep neural networks underpin a growing number of scientific and commercial applications. The broader field of generative AI, which relies on deep neural networks, has seen a significant increase in prevalence since the 2020s, with applications ranging from chatbots to text-to-image models [3]. Ensuring the reliability of these networks against adversarial perturbations remains an active area of research, particularly in high-stakes scientific domains where misclassifications can alter physics conclusions.