Plausible Adversarial Attacks on Direct Parameter Inference Models in Astrophysics by Benjamin Horowitz et al. on Tuesday 29 NovemberIn this abstract we explore the possibility of introducing biases in physicalparameter inference models from adversarial-type attacks. In particular, weinject small amplitude systematics into inputs to a mixture density networkstasked with inferring cosmological parameters from observed data. Thesystematics are constructed analogously to white-box adversarial attacks. Wefind that the analysis network can be tricked into spurious detection of newphysics in cases where standard cosmological estimators would be insensitive.This calls into question the robustness of such networks and their utility forreliably detecting new physics.arXiv: http://arxiv.org/abs/http://arxiv.org/abs/2211.14788v1