Account Takeover (ATO) fraud is an escalating threat in online marketplaces, but research progress remains limited due to the absence of domain-specific, publicly available datasets. This gap hinders benchmarking and reproducibility, slows methodological innovation, and prevents systematic comparisons between classical probabilistic models and modern deep learning approaches. This paper proposes MAATO, a new marketplace ATO dataset and a conceptual roadmap toward robust ATO detection by outlining the design principles for new synthetic datasets, intended to emulate realistic behavioral patterns and fraud scenarios. In parallel, we introduce a benchmarking framework grounded in the CRISP DM methodology to support reproducible evaluation across model families, feature engineering strategies, and anomaly scoring paradigms. Instead of reporting empirical findings, this paper articulates hypotheses regarding (1) the relative importance of engineered behavioral features, (2) the comparative performance of classical vs. deep learning architectures, and (3) the scalability of per-user versus global detection models. The work aims to guide future empirical studies and establish a foundation for shared community standards.
Jurišić et al. (Thu,) studied this question.