Forschungsthemen
[MA] Learning Reference Attribute Grammars
Problem. There are several methods to learn context-free grammars, but very few to learn context-sensitive, in particular Attribute Grammars, or Reference Attribute Grammars (RAG). The method AUTOGRAM <1> has shown how to learn context-free grammars from traces to recognize attackers in web systems. Objective. This work should investigate and identify the feasibility of RAG learning from example sentences, traces, or event traces. To this end, an extensive literature review should be provided by the thesis author. A proof-of-concept implementation of a learning should be realized and investigated with regard to its flexibility. Finally, several evaluation criteria need to be defined by the thesis author and evaluated with regard to validity. This thesis shall examine how to learn Reference Attributed Grammars by an extension of the AUTOGRAM method. To this end, the ideas in <2> on the learning of EMAG can be employed, but a new design and method development is required because the EMAG learning method is not well documented. The method shall be implemented as an extension of the JastAdd Grammar tool for RAG <3>. The following items are possible tasks: 1. Analyze the literature on grammar induction (grammar learning) and analyze the difference between related works. 2. Design and implement a new grammar learning algorithm, either as an extension of Autogram, EMAG learning, or related methods. 3. As an evaluation, a Penetration Test for one of the department's web systems for student management (jExam or MoleWeb) shall be constructed. The goal is to learn an "Attacker Recognition RAG" and recognize malicious attacks with the grammar that are similar to the trace data set. Alternatively, the learning algorithm can be evaluated by a collaborative robotic use case.
Betreuer: Johannes Mey-:#-#:- René Schöne