Bachelor's/Master's Thesis: Removing Alignment by Injecting PEFT Adapters

The Fraunhofer-Gesellschaft (www.fraunhofer.com) is one of the world's leading organizations for application-oriented research. 74 institutes develop pioneering technologies for our economy and society – more precisely: 30 000 people from technology, science, administration and IT. They know: Anyone who comes to Fraunhofer wants to and can make a difference. For themselves, for us and for the markets of today and tomorrow.

Background and Motivation

Parameter-efficient fine-tuning (PEFT) adapters are an excellent tool for adding new capabilities to LLMs with only a small number of parameters. Thanks to their reduced capacity, adapters are lightweight and therefore quick to train and distribute, while still achieving performance comparable to full fine-tuning. Methods from mechanistic interpretability make it possible to identify specific parameters in LLMs that are responsible for particular capabilities of the network, such as alignment. The goal of this thesis is to deactivate these safety parameters in the final model in a targeted way using PEFT adapters, thereby removing the alignment.

Goal

The challenge is that, due to the lower capacity of the adapters, parameters can no longer be masked in a targeted way. The goal is therefore to steer the adapters so that a large share of the safety neurons is still modified and the alignment is effectively bypassed.

Results

The intended result of the thesis is a comparison of the effectiveness of adapters with different capacities and methods. In addition to a proof of concept, the thesis will examine which adapter configuration is most vulnerable to such an attack.

Note

For a bachelor’s thesis, the scope can be reduced by considering only LoRA adapters.

Be part of change

Research and implement novel machine learning approaches that improve the safety of LLMs
Critically evaluate your results
Present your results
Write a project report in the form of a thesis

What you contribute

Very good academic performance with an average grade of 2.2 or better (German grading scale)
Knowledge of machine learning, including training, inference and optimization of transformer architectures
Knowledge of AI safety is desirable
Good Python skills are essential and will be tested during the interview
Interest in science and current research projects

What we offer

Flexible working hours
Insights into the intersection of academic research and industrial application

We value and promote the diversity of our employees' skills and therefore welcome all applications – regardless of age, gender, nationality, ethnic and social origin, religion, ideology, disability, sexual orientation and identity. Severely disabled persons are given preference in the event of equal suitability. Our tasks are diverse and adaptable – for applicants with disabilities, we work together to find solutions that best promote their abilities.

With its focus on developing key technologies that are vital for the future and enabling the commercial utilization of this work by business and industry, Fraunhofer plays a central role in the innovation process. As a pioneer and catalyst for groundbreaking developments and scientific excellence, Fraunhofer helps shape society now and in the future.

Ready for a change? Then apply now and make a difference! Once we have received your online application, you will receive an automatic confirmation of receipt. We will then get back to you as soon as possible and let you know what happens next.

Fraunhofer Institute for Secure Information Technology SIT

www.sit.fraunhofer.de

Requisition Number: 86083 Application Deadline:

Job Segment: Neurology, Research, Healthcare

Apply now »