Masterarbeten
Resultat från vårt Master Thesis Program – där studenter skriver exjobb i verkliga miljöer och ger organisationer den kompetens som krävs för att ta sina AI-projekt till nästa nivå.
Arbeten 2026
Identifying LLM-Powered Cyber Attacks with Timing Analysis and Honeytoken-Based Deception
Vår masteruppsats undersökte hur LLM-drivna angripare kan identifieras och skiljas från traditionella botar och mänskliga angripare med hjälp av beteendeanalys, tidsanalys och 'honey-tokens'. Våra resultat visade på tydliga skillnader i tidsmönster, beteende och interaktion med honey-tokens, vilket visar att LLM-drivna angripare har mätbara egenskaper som kan användas för detektion. I takt med att LLM:er och agentisk AI blir allt bättre på att utföra autonoma uppgifter inom cybersäkerhet, ger vårt arbete organisationer och forskare praktiska metoder för att upptäcka och analysera denna nya typ av cyberhot.
Purpose
The aim of this thesis is to investigate whether LLM-powered attackers can be reliably distinguished from automated bots and human attackers in SSH honeypot environments, and to propose a detection framework that integrates network-level, session-level, and deception-based signals toward this end.
Resources
LLM Attacker Behavior and Designing Adaptive Honeypot Defences
I takt med att autonoma AI-agenter börjar automatisera cyberattacker i flera steg saknar säkerhetsteam praktiska data om hur dessa verktyg faktiskt beteende sig när de utforskar system. Med utgångspunkt i forskning inom Project Violet undersökte detta arbete hur man kan kartlägga angripares avsikter samt testade om dynamiskt anpassningsbara 'honeypots' i realtid faktiskt förbättrar försvaret. Våra resultat visar att extra algoritmisk komplexitet inte nödvändigtvis överträffar enkla, välförankrade lösningar. Det hjälper team att fatta pragmatiska och kostnadseffektiva beslut när de experimenterar med AI-driven 'threat intelligence'.
Purpose
The purpose of this thesis is to address the empirical gap in current understanding of LLM based cyber offence and defence. By leveraging Project Violet, a specialised research framework containing alarge scale dataset of interactions between autonomous LLM attackers and LLM powered honeypots, this work moves from theoretical risk assessment to data-driven characterisation of AI driven threats.
Resources
Exploiting Large Language Models: Output Prefix Attack
Businesses shipping or buying LLM products can apply our takeaways immediately: audit output-prefill and reasoning surfaces before rollout, add prefix-style attacks to their red-team playbook, and treat model-provider choice as a security decision.
Purpose
Many commercial Large Language Models let users edit model responses and their "reasoning". Our research shows this can lead to simple exploitation at near-zero cost to the attacker. Using our Output Prefix Attack that starts a harmful answer, we developed a testing framework on DeepSeek, Gemini, and Claude commercial APIs. The results show the increased impact when targeting specifically model's reasoning, with attack success rates increased up to 99% on one model.
Resources
- Not yet public, shared individually
- Paper under review for publication
Improving the Performance of Shallow Neural Networks Using Norm-Preserving Regularization
I developed a new method of regularization that improved performance of shallow neural networks; useful in applications requiring fast inference such as IoT and networking. The method works across multiple domains (visual, language, tabular) and does not add significant computational overhead.
Purpose
The thesis investigated a new method of improving performance of neural nets with few layers. These so-called Shallow Neural Nets (SNNs) appear in edge inference applications such as networking and IoT where fast inference times are required or limited computing resources are available. They also provide a very efficient way of conducting transfer learning of a big foundation model to a specific task by retraining the classification head. The new method produced improved accuracy in the three classification tasks tested (visual, NLP, tabular) and one of the two transfer learning tasks (visual). The technique is implemented as a loss term that is added to the ML training loop and does not add a significant computational overhead which makes it more feasible than other methods. More research could be conducted on modifications to the regularization method as well as applying it to deeper architectures.
Resources
Warping the Search Space: No-Regret Bayesian Optimization with Finite-Library Input-Warped Kernels for Multi-Agent System Design
The thesis shows that automated optimization can identify strong multi-agent system designs: in the main experiment, the best automatically selected workflow performed 29.4% better than a human-engineered baseline in repeated evaluations. For companies developing agentic AI systems, the approach offers a reusable way to reduce reliance on manual architecture tuning and direct expensive experiments toward more promising configurations, potentially saving engineering time and compute while improving system performance.
Purpose
Develop and evaluate a more systematic, sample-efficient way to optimize multi-agent AI systems, where testing different agent configurations can be costly and manual trial-and-error does not scale well. The work explores how Bayesian optimization can search across choices such as agents, prompts, tools and workflow structure while making better use of a limited experimentation budget.
Resources
Fler masterarbeten publiceras inom kort.
