Fraudulent corporate and financial structures often rely on interposed agents and low-capacity entities controlled by a few ultimate beneficiaries, forming networks designed for asset shielding and concealment. This study analyzes graphs constructed from corporate, economic, and asset relationships to identify discriminative patterns between fraud and non-fraud cases. Three complementary dimensions are examined: (i) scale, through cardinality metrics (n, m); (ii) intensity and aggregation, through vertex weights ( \(w_i\) ), edge weights ( \(w_{i,j}\) ), and the mean weighted sum (S); and (iii) structural organization, through density ( \(\rho \) ), clustering (C), average shortest path length ( \(\langle d \rangle \) ), diameter (D), assortativity (r), and transitivity (T). Fraudulent networks tend to be smaller, more compact, and more heterogeneous in vertex and edge weights, with higher local density and redundant ties, whereas non-fraudulent networks are larger, more diffuse, and less centralized. Interpreted from the perspective of the Brazilian Office of the Attorney General of the National Treasury (PGFN), these findings provide robust structural signals of fraudulent arrangements and practical support for investigative prioritization, credit recovery, and risk monitoring. Future research should extend this framework to applied machine learning pipelines for supervised classification in real-world tax enforcement scenarios.