A Behavioural Bot Detection Framework Based on the Dirichlet-Multinomial Log-Likelihood

This thesis addresses the challenge of distinguishing human users from automated bots in web traffic, an increasingly relevant problem with the rise of LLM-based agents. While traditional bot detection techniques usually focus on network-level or interaction features, often failing against advanced bots that mimic human behaviour, this work focuses instead on the analysis of thematic patterns in user sessions, proposing a behavioural bot detection framework. User activity is modelled using the Dirichlet-Multinomial distribution, which captures the degree of heterogeneity (known as overdispersion) in accessed content. The main hypothesis is that human browsing behaviour is thematically dispersed, while that of bots is more concentrated. The proposed method uses the Dirichlet-Multinomial log-likelihood function and its associated parameters as a detection score. The framework is built upon recent advances that make the computation of the model efficient and stable on large datasets, enabling its practical application to real-world traffic analysis.