Theses

The project concept “AI-based analysis and utilization of physical signal parameters for securing smart home systems” involves researching and prototypically developing the so-called “Physical-Guards” approach. This contributes to the research area of “Physical Layer Security” by adapting and further developing the state of science and technology from the fields of IT security and data science/machine learning. The foundation for the research approach is the exploitation of physical parameters (e.g., local signal strength) of wireless communication in smart homes to a) detect attacks on the smart home infrastructure, b) secure the integrity of signals, and c) ensure the confidentiality of security-critical information. Compared to or in conjunction with other IT security measures, Physical-Guards technology promises the following advantages: the physical properties of communication signals are subject to laws of nature. Considering these properties through security infrastructure makes many attack strategies impossible or significantly more difficult. Additionally, security concepts based on the physical layer have the advantage that in smart homes, the large heterogeneity of standards and communication protocols at higher layers does not limit the deployability of security technology due to technical incompatibilities with IoT devices. The project consortium consisting of the University of Mannheim (specifically the two research groups Dependable Systems Engineering (IT security) and Institute for Enterprise Systems (AI/Machine Learning)) as well as the medium-sized industrial technology partners M2M Germany and osapiens.

Goals or topics of a possible thesis could be:

1. detect attacks on the smart home infrastructure,
2. secure the integrity of signals, and/or
3. ensure the confidentiality of security-critical information.

As these areas are rather broad, a potential thesis topic will be a subset of the above stated.

For further details, questions and/or proposals of your own ideas connected with the descripted project, please feel free to contact Yves T. Staudenmaier.

Many modern Machine Learning (ML) systems rely on collecting, aggregating, and processing sensitive data (e.g., health records, location traces, user behavior). Privacy-Preserving Machine Learning (PPML) studies how to train and deploy useful models while minimizing what is revealed about individuals, datasets, or model internals, even in the presence of curious (or malicious) parties. In security and privacy research, PPML can be approached from two complementary angles:

• Privacy as a requirement: How can we train/infer without exposing sensitive inputs or leaking training data?
• Privacy as an attack surface: What can an adversary infer from model outputs, gradients, updates, or access patterns, and how do we mitigate it?

Research topics in this area include (examples):

Privacy Attacks on ML

• Membership Inference: Determining whether a specific record was part of the training set
• Attribute Inference: Inferring sensitive attributes about data subjects from model behavior
• Model Inversion / Reconstruction: Reconstructing likely training inputs from outputs or intermediate signals

Defenses and PPML Techniques

• Differential Privacy (DP): DP-SGD and privacy accounting; utility–privacy trade-offs; auditing privacy claims
• Federated Learning (FL): Secure aggregation, robustness vs. privacy, client dropouts, cross-device vs. cross-silo settings
• Secure Multi-Party Computation (MPC): Training or inference via secret sharing; performance/accuracy trade-offs
• Homomorphic Encryption (HE): Encrypted inference and (limited) training; model design for HE-friendly computation

PPML for Practical Security Settings

• Evaluation frameworks: Benchmarking PPML schemes for latency, accuracy, privacy guarantees, and threat models

If one of these topics sparks your interest, or if you have an idea of your own, just contact Marcel Mildenberger. Please note that you should have prior experience in ML (PyTorch/TensorFlow) and be comfortable reading current research papers. 

Depending on the topic, familiarity with at least one of the following is helpful: differential privacy, federated learning, cryptography (MPC/HE), or system security/threat modeling.

Bitcoin and similar cryptocurrencies are a great opportunity for research. Possible topics include:

• Learning vector respresentations of blockchain entities
• Forensic analysis of cryptocurrency services
  Many crytocurrency services, like exchanges, marketplaces, casinos etc., leave distinctive transaction patterns on the blockchain. Your task would be to identify such patterns in order to gain insights on the operations of these services.
• Development of a cryptocurrency forensics tool
  Your task would be to design and implement the prototype of a cryptocurrency forensics tool. The tool should provide a GUI allowing users to look up Bitcoin addresses, as well as a convenient way of addig new identified addresses to the underlying database.

If you're interested in writing a thesis on one of these topics or you want to discuss your own research idea, just contact Jochen Schäfer.

In the context of security research, machine learning can be seen from two angles: Either as a possible target or as a tool for carrying out/defending against attacks. For example, there has been research in the following areas:

• Attacks on Machine Learning
  • Data Extraction: Recovering training data from a trained model
  • Model Extraction: Recovering model parameters from a trained model
  • Data Poisoning: Injecting malicious traning data to manipulate the classifier
• ML Applications in Security Research
  • Code/Instruction embeddings for malware classification
  • Identifying code authors based on coding style

If one of these topics sparks your interest, or if you have an idea on your own, just contact Jochen Schäfer. Please note that you should have some prior experience in the ML domain, particularly with the models you intend to use.

Description:
CrypTool 2 is an open-source program that allows you to try out various cryptographic methods. CrypTool 2 provides a visual programming interface which easily can be used to integrate and manipulate cryptographic functions into workflows. More specifically, the individual cryptographic methods are implemented by so-called plug-ins, which are represented by individual graphical objects. These can be combined with drag & drop on the graphical user interface. This approach makes it easy to visualise complex processes and thus to understand them better.

Goal: 
Several existing cryptographic methods are to be implemented so that they can be officially recorded in CrypTool 2. This also includes the creation of documentation, the clean structuring and commenting of the source code, etc. The exact selection of the topics to be implemented is discussed with the student and the CrypTool 2 team.

Contact person: Frederik Armknecht

Privacy-Preserving Record-Linkage (PPRL) techniques have been developed to link persons  without violating their privacy. However, Graph Matching Attacks allow the re-identification of individuals in the encoded database. These Graph Matching Attacks are currently seen as the most serious threat to many PPRL schemes. In your thesis, you may work on one of the following topics:

• Development of new and improved (Graph Matching) Attacks against PPRL.
• Development of countermeasures to protect PPRL schemes against existing attacks.

Contact person: Jochen Schäfer

After agreeing on a topic, we discuss the expected content and agree on a preliminary structure, usually in the form of a table of contents. This may serve you as a guide while writing your thesis.

Before you can start working on your thesis, we need additional information for the official thesis registration. The required information consists of

• Your full name
• Your matriculation number
• Your address
• Your pursued degree
• Your study program
• The language of writing (i.e. English or German)

The deadline for submitting your thesis is set with registration and depends on the type of thesis and your examination regulations – usually, you have three months for working on a Bachelor's thesis and six months for a Master's thesis.

Unlike other assessments, a thesis can be started flexibly during a semester irrespective of lecture or examination periods.

It is your responsibility to organize and manage the time available for writing your thesis. Formally, you are not required to keep your supervisor(s) up to date. However, we recommend contacting your supervisor(s) regularly as it helps you to stay on track and to tackle potentially emerging problems early on. Discuss your preferred style of supervision with your supervisor(s).

You have to submit your thesis no later than the deadline as defined by the registration. As per most examination regulations, you are required to hand in two printed copies including a signed affidavit (cf. your examination regulation for the exact wording), and a digital copy, preferably as a PDF file. You should send the digital version directly to your supervisor(s). Consider using duplex printing for the two hard copies with a binding of your choice.

Depending on your examination regulations, you may have to present your thesis in form of a talk. However, even if you are not obliged to, we encourage everyone to seize this opportunity as you can strengthen your presentation skills, especially when presenting scientific work to an academic and diverse audience. Such a talk should take about 30 minutes and include five minutes for questions from the audience. Please note that any voluntarily given talk will not be graded.