This thesis stems from the need to validate the framework for detecting bot-generated web traffic proposed in the thesis ``A Behavioural Bot Detection Framework based on the Dirichlet Multinomial Log-Likelihood Function'' by Genesin Silvio (A.Y. 2025/2026), based on the Dirichlet-Multinomial distribution and its log-likelihood to distinguish the behaviour of real users from that of automated systems. Genesin's work defines its theoretical foundation, while this thesis aims to experimentally answer two questions: does the framework actually work, and is the LM (Languasco-Migliardi) procedure actually necessary for the numerically stable computation of the log-likelihood? To work around the absence of a reliably labelled real-world dataset (the ground truth problem), a synthetic and controlled testbed is built by sampling sessions directly from the Dirichlet-Multinomial generative process assumed by the framework; the implementation relies on the PyPI package alm2, which implements the LM procedure. On the entire IAB taxonomy of 703 categories, the classifier achieves perfect separation between human and bot sessions (accuracy, precision, recall, F1 and AUC all equal to 1.0), with a gradual degradation of discriminative power as the separation between the generative classes decreases. The comparison between direct computation and the LM procedure shows that the two techniques agree to more than 7 significant figures up to 1/ψ ≈ 10^8-10^9, beyond which the direct computation diverges, confirming the numerical necessity of the LM procedure as the dimensionality of the problem grows. The work thus provides an experimental verification of the applicability of the framework and a basis for subsequent validation on real web traffic data.
Questa tesi nasce dalla necessità di validare il framework per il riconoscimento del traffico web generato da bot proposto nella tesi “A Behavioural Bot Detection Framework based on the Dirichlet Multinomial Log-Likelihood Function” di Genesin Silvio (A.A. 2025/2026), basato sulla distribuzione Dirichlet-Multinomiale e sulla relativa log-verosimiglianza per distinguere il comportamento di utenti reali da quello di sistemi automatizzati. Il lavoro di Genesin ne definisce l'impostazione teorica, mentre questa tesi si pone l'obiettivo di rispondere sperimentalmente a due domande: il framework funziona davvero, ed è davvero necessaria la procedura LM (Languasco-Migliardi) per il calcolo numericamente stabile della log-verosimiglianza? Per aggirare l'assenza di un dataset reale affidabilmente etichettato (ground truth problem), viene costruito un banco di prova sintetico e controllato, campionando le sessioni direttamente dal processo generativo Dirichlet-Multinomiale assunto dal framework; l'implementazione si basa sul pacchetto PyPI alm2, che realizza la procedura LM. Sull'intera tassonomia IAB di 703 categorie il classificatore ottiene una separazione perfetta fra sessioni umane e bot (accuratezza, precisione, recall, F1 e AUC pari a 1.0), con un degrado graduale del potere discriminante al ridursi della separazione tra le classi generative. Il confronto tra calcolo diretto e procedura LM mostra che le due tecniche concordano su più di 7 cifre significative fino a 1/ψ ≈ 10^8-10^9, oltre le quali il calcolo diretto diverge, confermando la necessità numerica della procedura LM al crescere della dimensionalità. Il lavoro fornisce così una verifica sperimentale dell'applicabilità del framework e una base per successive validazioni su dati di traffico web reali.
Behavioural Bot Detection tramite la Distribuzione Dirichlet-Multinomiale: Implementazione e Validazione della Procedura LM su dati sintetici
BOSCOLO MENEGUOLO, MATTIA
2025/2026
Abstract
This thesis stems from the need to validate the framework for detecting bot-generated web traffic proposed in the thesis ``A Behavioural Bot Detection Framework based on the Dirichlet Multinomial Log-Likelihood Function'' by Genesin Silvio (A.Y. 2025/2026), based on the Dirichlet-Multinomial distribution and its log-likelihood to distinguish the behaviour of real users from that of automated systems. Genesin's work defines its theoretical foundation, while this thesis aims to experimentally answer two questions: does the framework actually work, and is the LM (Languasco-Migliardi) procedure actually necessary for the numerically stable computation of the log-likelihood? To work around the absence of a reliably labelled real-world dataset (the ground truth problem), a synthetic and controlled testbed is built by sampling sessions directly from the Dirichlet-Multinomial generative process assumed by the framework; the implementation relies on the PyPI package alm2, which implements the LM procedure. On the entire IAB taxonomy of 703 categories, the classifier achieves perfect separation between human and bot sessions (accuracy, precision, recall, F1 and AUC all equal to 1.0), with a gradual degradation of discriminative power as the separation between the generative classes decreases. The comparison between direct computation and the LM procedure shows that the two techniques agree to more than 7 significant figures up to 1/ψ ≈ 10^8-10^9, beyond which the direct computation diverges, confirming the numerical necessity of the LM procedure as the dimensionality of the problem grows. The work thus provides an experimental verification of the applicability of the framework and a basis for subsequent validation on real web traffic data.| File | Dimensione | Formato | |
|---|---|---|---|
|
Boscolo_Meneguolo_Mattia.pdf
accesso aperto
Dimensione
941.1 kB
Formato
Adobe PDF
|
941.1 kB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/120214