Theses

If you are interested in writing your thesis at our chair, please review our available topics. If you are interested in a specific topic, please contact the corresponding assistant.

If you are interested in multiple research areas at the chair, please let us know. We will gladly help you choose a suitable topic.

 

TopicTypeAssistantStatus
Understanding and analysing safety alignment in LLMsB.Sc. / M.Sc. ThesisHamid DashtbaniReserved
Analysing representations of harmful concepts in LLMsB.Sc. / M.Sc. ThesisHamid DashtbaniAvailable
Protecting privacy in LLMs with homomorphic encryptionB.Sc. / M.Sc. ThesisHamid DashtbaniAvailable
Facilitating proof of work in cryptocurrencies with LLMsB.Sc. / M.Sc. ThesisHamid DashtbaniAvailable
Are safety neurons identifiable? Non-uniqueness of safety mechanisms in LLMsB.Sc. / M.Sc. ThesisAnsgar SchäfftleinAvailable
Comparing neuron-, weight-, and subspace-level vulnerabilities in safety aligned LLMsB.Sc. / M.Sc. ThesisAnsgar SchäfftleinAvailable
Improving safety alignment training through safety neuron splittingB.Sc. / M.Sc. ThesisAnsgar SchäfftleinAvailable
Defending against low-rank malicious fine-tuning with high-rank safety alignmentB.Sc. / M.Sc. ThesisAnsgar SchäfftleinAvailable
Dynamic analysis for automated vulnerability repairB.Sc. / M.Sc. ThesisVarun GadeyReserved
White-box mitigation of vulnerabilities in AI-generated codeB.Sc. / M.Sc. ThesisVarun GadeyReserved
Secure code generation for mixture-of-experts LLMsB.Sc. / M.Sc. ThesisVarun GadeyAvailable
Mitigating hashjacking attacks on AI browsers using MLB.Sc. / M.Sc. ThesisDr. Venkata MarellaAvailable
Memory-Constrained Neural Architecture Search for On-Device Deepfake Audio Detection on ESP32B.Sc. / M.Sc. ThesisBehrad BayatReserved
Secure and robust model watermarking in federated learningB.Sc. / M.Sc. ThesisBehrad BayatReserved
Cross-client memory backdoor attacks in federated learningB.Sc. / M.Sc. ThesisBehrad BayatReserved