Master's Thesis: Attacking Current Backdoor Detection Methods

Fraunhofer-Gesellschaft · Darmstadt

ThesisGerman needed: not said

Still open on 1 Nov

You apply on the company's own website.

Key facts

Background/Motivation: Backdoor attacks are attacks on neural networks where a so-called trigger changes the decision-making behavior of the networks and thereby creates vulnerabilities. These triggers can be smuggled into the training dataset or directly into the model weights. These are then called poisoned. Through parameter-efficient fine-tuning methods, backdoor attacks on Large Language Models (LLMs) have become significantly more difficult to detect, since a poisoned parameter update is harder to recognize than a poisoned dataset. Therefore, several methods have been developed recently to detect poisoned model updates.
Goal: Due to the diversity of backdoor attacks, methods often detect far fewer attacks than they claim, because they often make assumptions that do not correspond to reality. The goal of this thesis is therefore to find and exploit vulnerabilities in methods presented in the literature, so that promised effects are not achieved as desired.
Results: The results of this thesis should show the research community that a fundamental understanding of the mechanics of backdoor attacks is essential. For this purpose, models, datasets or fine-tuning adapters shall be created and tested for previously selected detection methods that can reliably bypass these detection methods.

The company did not list what you need. Open the job on the company site to read it.

Background/Motivation: Backdoor attacks are attacks on neural networks where a so-called trigger changes the decision-making behavior of the networks and thereby creates vulnerabilities. These triggers can be smuggled into the training dataset or directly into the model weights. These are then called poisoned. Through parameter-efficient fine-tuning methods, backdoor attacks on Large Language Models (LLMs) have become significantly more difficult to detect, since a poisoned parameter update is harder to recognize than a poisoned dataset. Therefore, several methods have been developed recently to detect poisoned model updates.
Goal: Due to the diversity of backdoor attacks, methods often detect far fewer attacks than they claim, because they often make assumptions that do not correspond to reality. The goal of this thesis is therefore to find and exploit vulnerabilities in methods presented in the literature, so that promised effects are not achieved as desired.
Results: The results of this thesis should show the research community that a fundamental understanding of the mechanics of backdoor attacks is essential. For this purpose, models, datasets or fine-tuning adapters shall be created and tested for previously selected detection methods that can reliably bypass these detection methods.

The company did not list what you need. Open the job on the company site to read it.