Offered Theses

Identification is a communication scheme that allows rate doubly exponential in the blocklemght, with the tradeoff that identities cannot be decoded (as messages do) but can only be verified.

• https://ieeexplore.ieee.org/document/42172
• https://ieeexplore.ieee.org/document/42173

Identification codes can be achieved by first removing the errors from the channel with regular transmission channel coding, and then sending a challenge though the corrected channel. For every identity i, The channenge is generated by picking a random input m and computing the corresponding output T_i(m) using a function T_i that depends on the identity. The challenge is then the pair m,T_i(m) and the receiver wanting to verify an identity j will verify whether j=i by testing the challenge. This is done by recomputing the output with T_j and verifying whether T_j(m)= T_i(m). The errors are reduced by ensuring that the various functions collide on a small fraction of the possible inputs.

It turns out that choosing good sets of funtions {T_i} is the same as choosing error-correction codes {c_i} with large distance, where now each codeword c_i defines a function by mapping positions m (sometimes called code locators) to symbols c_im of the codeword.
We can thus construct identification codes by choosing error-correction codes where we are only interested in the performance of the error correction encoders (we are not interested in the error-correction decoder or error-correction codes).

From previous work we have a fairly efficient implementation based Reed-Muller code which can be found at

• https://arxiv.org/abs/2107.07649
• https://arxiv.org/abs/2107.06801
• https://arxiv.org/abs/2007.06372

Secrecy in this identification codes has also been implemented in unpublished work. Furthermore, the theoretical work on Identification with PUF's has been done in

• https://ieeexplore.ieee.org/document/8445910
• https://ieeexplore.ieee.org/document/8600405

The goal of the project will be to bridge the three topics and create practical and efficient secret identification codes in the PUF setting.

The working language will be in English.

Environment: this is a project in collaboration with LTI. At LNT and LTI there is currently a lot of funding for research in identification. Therefore you will find a large group of people that might be available for discussion and collaboration.

For different applications, the demand for reliable memory solutions in particular for non-volatile memories such as phase-change memories (PCMs)  is rapidly increasing. PCM cells may become defective (also called stuck) either fully or partially if they fail in switching their states, and therefore these cells can only hold a single phase. In response to these defects, combined masking and error-correcting code constructions have been proposed, where masking is for hiding the defects while error-correcting is to compromise potential added-channel errors.  We want to investigate further code constructions such that less overall redundancy is required to handle these two types of errors. As an alternate, work for combined erasure errors and masking code constructions could be investigated.

- Basic principle of Linear Algebra 

- Channel Coding/Coding Theory 

- Basic knowledge in Information Theory  

Identification is a communication scheme that allows rate doubly exponential in the blocklemght, with the tradeoff that identities cannot be decoded (as messages do) but can only be verified.

• https://ieeexplore.ieee.org/document/42172
• https://ieeexplore.ieee.org/document/42173

The double exponential growth presents various challenges in the finite regime: there are heavy computational costs introduced at the encoder and decoder and heavy trade-offs between the error and the codes sizes.

The ultimate goal is to find a fast, reliable implementation while still achieving large code sizes.

Identification codes can be achieved by first removing the errors from the channel with regular transmission channel coding, and then sending a challenge though the corrected channel. For every identity i, The channenge is generated by picking a random input m and computing the corresponding output T_i(m) using a function T_i that depends on the identity. The challenge is then the pair m,T_i(m) and the receiver wanting to verify an identity j will verify whether j=i by testing the challenge. This is done by recomputing the output with T_j and verifying whether T_j(m)= T_i(m). The errors are reduced by ensuring that the various functions collide on a small fraction of the possible inputs.

It turns out that choosing good sets of funtions {T_i} is the same as choosing error-correction codes {c_i} with large distance, where now each codeword c_i defines a function by mapping positions m (sometimes called code locators) to symbols c_im of the codeword.
We can thus construct identification codes by choosing error-correction codes where we are only interested in the performance of the error correction encoders (we are not interested in the error-correction decoder or error-correction codes).

Your task will be implementing the identification codes described in

• https://ieeexplore.ieee.org/abstract/document/782144

aiming at the fastest implementation, and testing their performance in comparison to other current implementations.

For reference, our previous work on identification based on Reed-Solomon and Reed-Muller code can be found at

• https://arxiv.org/abs/2107.07649
• https://arxiv.org/abs/2107.06801
• https://arxiv.org/abs/2007.06372

The coding will be in Python/Sagemath.
The working language will be in English.

Environment: we collaborate with LTI. At LNT and LTI there is currently a lot of funding for research in identification. Therefore you will find a large group of people that might be available for discussion and collaboration.

In this project, we investigate the interplay between redundancy and straggler tolerance in distributed learning.

The setting is that of a main node distributing computational tasks to available workers as part of a machine learning algorithm, e.g., training a neural network. Waiting for all workers to return their computations suffers from the presence of stragglers, i.e., slow or unresponsive nodes. Mitigating the effect of the stragglers can be done through the use of redundancy or by leveraging the properties of the convergence of the machine learning algorithm. 

The goal of this work is to compare when redundancy is needed. In this case, we aim to first analyze the convergence speed with and without redundancy. Then, we aim at design schemes that adaptively increase the redundancy to speed up the convergence.

Further reading: 

R. Bitar, M. Wootters and S. El Rouayheb, Stochastic Gradient Coding for Straggler Mitigation in Distributed Learning, IEEE Journal on Selected Areas in Information Theory (JSAIT), Vol. 1, No. 1, May 2020. arXiv:1905.05383

S. Kas Hanna, R. Bitar, P. Parag, V. Dasari and S. El Rouayheb, Adaptive Stochastic Gradient Descent for Fast and Communication-Efficient Distributed Learning, preprint, arXiv:2208.03134.

Knowledge in the following topics:

• Probability Theory
• Gradient descent and stochastic gradient descent
• Coding theory

Independence and motivation to work on a research topic

Knowledge of implementing neural networks is a plus

Space-division multiplexing (SDM), which consists in exploiting multimode fibers (MMFs) or multicore fibers (MCFs) instead of single mode ones, is one of the future optical communications architectures to increase data rates and network planning flexibility. The nonlinear properties of MCFs are of primary interest in assessing the usefulness of SDM against the current network. With this thesis, the student has the chance to work on a state-of-the-art topic in the field of optical communication systems, and progress quickly thanks to a tight (if desired) supervision. Would you be curious to know more about it? If so, just get in touch with me at paolo.carniello@tum.de (personal page https://www.ce.cit.tum.de/lnt/mitarbeiter/doktoranden/carniello/).

-some knowledge on optical communications systems (e.g., Optical Communication Systems or Simulation of Optical Communication Systems Lab)

-some knowledge about communications engineering topics

Asymmetric cryptography is a core component of modern communication infrastructure. The existence of a sufficiently large quantum computer threatens all algorithms currently in use and recent developments in this field motivate the field of post-quantum cryptography (PQC), with the primary objective of developing secure and futureproof alternatives. To consolidate these efforts, the National Institute of Standards and Technology (NIST) is conducting a competition with the goal of selecting and standardizing the best available candidates.

In July 2022 four algorithms were selected, one Public-key Encryption and Key-establishment Algorithms (KEM) and three Digital Signature Algorithms (DSA). Of these four algorithms, three (including the two recommended for general purpose applications) are from the class of lattice-based schemes, i.e., they rely on difficult problems over lattices for their security.

The lattices in these schemes are represented by elements from a polynomial ring and arithmetic over this ring therefore plays a crucial role in their execution. To accelerate this arithmetic and, specifically, the multiplication of polynomials, the number-theoretic transform (NTT) is used. This approach is a generalization of the multiplication algorithms based on the fast Fourier transform (FFT), that have been long established in fields like signal processing. Since a considerable part of the computational complexity of these algorithms lies in this NTT, it is a prime candidate for HW acceleration and many works in literature have proposed such accelerators.

While considerable efforts have been made to offer fast and lean NTT accelerators, the topic of fault resilience has received little attention so far. In safety critical applications, as common in automotive or industrial fields, this resilience is an important feature and needs to be provided by the HW. On the other hand, these fields are traditionally price sensitive, so the additional chip area required for these features should be minimal. However, current approaches, such as those introduced by Sarker et al. [1], impose a large area (or latency) overhead.

The goal of this thesis is to address this shortcoming. First, the approaches published in literature shall be evaluated with regard to the relevant performance figures (area overhead, latency, …). Then, new approaches for error resilience in NTT calculation shall be developed, either based on improving upon existing methods or by adapting methods for error resilience in the computation of FFTs.

References

[1] Sarker, Ausmita, Alvaro Cintas Canto, Mehran Mozaffari Kermani, and Reza Azarderakhsh. “Error Detection Architectures for Hardware/Software Co-Design Approaches of Number-Theoretic Transform.” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2022, 1–1. https://doi.org/10.1109/TCAD.2022.3218614.

Basic understanding of HW design

Good knowledge of algebra

Cryptocurrencies like Bitcoin and Ethereum use a decentralized ledger called Blockchain to track transactions. Whenever a new block is added to the Blockchain, the change is spread through the network using a gossip-like protocol. This process is known as block propagation.

To increase scalability, the efficiency of block propagation is crucial. This thesis aims to explore the information theoretic limits of block propagation, derive realistic models based on real data, and investigate innovative and efficient techniques for block propagation.

The thesis will be conducted at the Institute of Communications and Navigation at DLR (German Aerospace Center) in Oberpfaffenhofen.

Required qualifications are

• basic knowledge of information theory
• programming experience in Matlab, C, or python.
• Interest in cryptocurrencies.

Identification is a communication scheme that allows rate doubly exponential in the blocklemght, with the tradeoff that identities cannot be decoded (as messages do) but can only be verified.

• https://ieeexplore.ieee.org/document/42172
• https://ieeexplore.ieee.org/document/42173

Identification codes can be achieved by first removing the errors from the channel with regular transmission channel coding, and then sending a challenge though the corrected channel. For every identity i, The channenge is generated by picking a random input m and computing the corresponding output T_i(m) using a function T_i that depends on the identity. The challenge is then the pair m,T_i(m) and the receiver wanting to verify an identity j will verify whether j=i by testing the challenge. This is done by recomputing the output with T_j and verifying whether T_j(m)= T_i(m). The errors are reduced by ensuring that the various functions collide on a small fraction of the possible inputs.

It turns out that choosing good sets of funtions {T_i} is the same as choosing error-correction codes {c_i} with large distance, where now each codeword c_i defines a function by mapping positions m (sometimes called code locators) to symbols c_im of the codeword.
We can thus construct identification codes by choosing error-correction codes where we are only interested in the performance of the error correction encoders (we are not interested in the error-correction decoder or error-correction codes).

From previous work we have a fairly efficient implementation based Reed-Muller code which can be found at

• https://arxiv.org/abs/2107.07649
• https://arxiv.org/abs/2107.06801
• https://arxiv.org/abs/2007.06372

Secrecy in this identification codes has also been implemented in unpublished work. Furthermore, the theoretical work on Identification with PUF's has been done in

• https://ieeexplore.ieee.org/document/8445910
• https://ieeexplore.ieee.org/document/8600405

The goal of the project will be to bridge the three topics and create practical and efficient secret identification codes in the PUF setting.

The working language will be in English.

Environment: this is a project in collaboration with LTI. At LNT and LTI there is currently a lot of funding for research in identification. Therefore you will find a large group of people that might be available for discussion and collaboration.

As part of an ongoing project with Huawei we are looking into quantum machine learning algorithms applied to decoding at the end of an optical fiber in the non-linear regime.

So far we have tried only the quantum version of k-mean clustering, however the goal is to test further quantum algorithms, in particular quantum support vector machines next, and their classical quantum-inspired counterpart.

The projects will involve reading the literature on quantum machine learning algorithms and quantum-inspired algorithms, find or come up with an implementation (this will involve the use of quantum libraries, in particular so far we have use qiskit), and benchmark the performance.

Knowledge of quantum mechanics or quantum information is highly recommended.

For different applications, the demand for reliable memory solutions in particular for non-volatile memories such as phase-change memories (PCMs)  is rapidly increasing. PCM cells may become defective (also called stuck) either fully or partially if they fail in switching their states, and therefore these cells can only hold a single phase. In response to these defects, combined masking and error-correcting code constructions have been proposed, where masking is for hiding the defects while error-correcting is to compromise potential added-channel errors.  We want to investigate further code constructions such that less overall redundancy is required to handle these two types of errors. As an alternate, work for combined erasure errors and masking code constructions could be investigated.

- Basic principle of Linear Algebra 

- Channel Coding/Coding Theory 

- Basic knowledge in Information Theory  

Identification is a communication scheme that allows rate doubly exponential in the blocklemght, with the tradeoff that identities cannot be decoded (as messages do) but can only be verified.

• https://ieeexplore.ieee.org/document/42172
• https://ieeexplore.ieee.org/document/42173

The double exponential growth presents various challenges in the finite regime: there are heavy computational costs introduced at the encoder and decoder and heavy trade-offs between the error and the codes sizes.

The ultimate goal is to find a fast, reliable implementation while still achieving large code sizes.

Identification codes can be achieved by first removing the errors from the channel with regular transmission channel coding, and then sending a challenge though the corrected channel. For every identity i, The channenge is generated by picking a random input m and computing the corresponding output T_i(m) using a function T_i that depends on the identity. The challenge is then the pair m,T_i(m) and the receiver wanting to verify an identity j will verify whether j=i by testing the challenge. This is done by recomputing the output with T_j and verifying whether T_j(m)= T_i(m). The errors are reduced by ensuring that the various functions collide on a small fraction of the possible inputs.

It turns out that choosing good sets of funtions {T_i} is the same as choosing error-correction codes {c_i} with large distance, where now each codeword c_i defines a function by mapping positions m (sometimes called code locators) to symbols c_im of the codeword.
We can thus construct identification codes by choosing error-correction codes where we are only interested in the performance of the error correction encoders (we are not interested in the error-correction decoder or error-correction codes).

Your task will be implementing the identification codes described in

• https://ieeexplore.ieee.org/abstract/document/782144

aiming at the fastest implementation, and testing their performance in comparison to other current implementations.

For reference, our previous work on identification based on Reed-Solomon and Reed-Muller code can be found at

• https://arxiv.org/abs/2107.07649
• https://arxiv.org/abs/2107.06801
• https://arxiv.org/abs/2007.06372

The coding will be in Python/Sagemath.
The working language will be in English.

Environment: we collaborate with LTI. At LNT and LTI there is currently a lot of funding for research in identification. Therefore you will find a large group of people that might be available for discussion and collaboration.

As part of the NewCom Project, new communication paradigms are investigated from an experimental perspective in order to construct proof-of-concept implementations that demonstrate the theoretical results obtained for Post-Shannon Communication schemes. In particular, this MSc thesis focuses on Identification Codes and their integration into a simulation environment where vehicular networks are modelled.

For this, the master student will first conduct a review of the state-of-the-art use cases for identification in the scientific literature and in form of patents, with an emphasis on V2X communications. By using an open-source V2X implementation based on LDR’s Simulation of Urban Mobility (SUMO) framework integrated with ns-3’s implementation of the ITS-G5 and LTE standards and conducting simulation in specific scenarios, the student will gain a first impression of the performance of the system using traditional transmission schemes. The integration of existing implementation of identification codes culminates this thesis, where KPIs will be defined in order to compare the advantages of using identification instead of transmission in the context of V2X communications.

Details about the C++ tools/libraries

The software used for the simulation of the vehicular network communication is ezCar2x

• https://www.ezcar2x.fraunhofer.de/en.html
• https://gitlab.cc-asp.fraunhofer.de/ezcar2x

which build on and integrates the NS-3 (network simulation) and SUMO (traffic simulation) libraries

• https://sumo.dlr.de/docs/index.html
• https://www.nsnam.org/

For the identification part and identification code based on Reed-Muller codes needs to be reimplemented (work in progress) from Python into C++ using the Givaro library

• https://github.com/linbox-team/givaro
• https://casys.gricad-pages.univ-grenoble-alpes.fr/givaro/

• Knowledge of communications engineering, mobile communications, wireless channel models, signal processing, and channel coding techniques (experience in LTE/5G cellular networks is a plus)

• Interest in novel communication concepts as well in their practical implementation

• Software experience: MATLAB, C++ and Python (experience with ns-3 or SUMO is a plus)

• Comfortable working with Linux operative systems and distributed version control tools (e.g., gitlab)

• Goal-oriented and structured work style

Identification is a communication scheme that allows rate doubly exponential in the blocklemght, with the tradeoff that identities cannot be decoded (as messages do) but can only be verified.

• https://ieeexplore.ieee.org/document/42172
• https://ieeexplore.ieee.org/document/42173

The double exponential growth presents various challenges in the finite regime: there are heavy computational costs introduced at the encoder and decoder and heavy trade-offs between the error and the codes sizes.

The ultimate goal is to find a fast, reliable implementation while still achieving large code sizes.

Identification codes can be achieved by first removing the errors from the channel with regular transmission channel coding, and then sending a challenge though the corrected channel. For every identity i, The channenge is generated by picking a random input m and computing the corresponding output T_i(m) using a function T_i that depends on the identity. The challenge is then the pair m,T_i(m) and the receiver wanting to verify an identity j will verify whether j=i by testing the challenge. This is done by recomputing the output with T_j and verifying whether T_j(m)= T_i(m). The errors are reduced by ensuring that the various functions collide on a small fraction of the possible inputs.

It turns out that choosing good sets of funtions {T_i} is the same as choosing error-correction codes {c_i} with large distance, where now each codeword c_i defines a function by mapping positions m (sometimes called code locators) to symbols c_im of the codeword.
We can thus construct identification codes by choosing error-correction codes where we are only interested in the performance of the error correction encoders (we are not interested in the error-correction decoder or error-correction codes).

Your task will be speeding up the current implementations based on Reed-Solomon and Reed-Muller codes:

• https://arxiv.org/abs/2107.07649
• https://arxiv.org/abs/2107.06801
• https://arxiv.org/abs/2007.06372

The coding will be in Python/Sagemath.
This work can accomodate multiple students.
The working language will be in English.

Environment: we collaborate with LTI. At LNT and LTI there is currently a lot of funding for research in identification. Therefore you will find a large group of people that might be available for discussion and collaboration.

Nachrichtentechnik 2

The goal of this project is to implement in Python/Sagemath the security functions (at least one of four) described in https://arxiv.org/abs/2102.00983
Sagemath contains libraries for mosaics, BIBDs, etc, that can be used for the project.

Motivation:
There are various types of security definitions.
The mutual information based types, in increasing order of security requirement are

1. Weak secresy asks that the average mutual information of the eavesdropper I(M:E)/n goes to 0 for a uniform message M (average here means averaged over the blocklength n, an additional average over M is implicit in the mutual information)
2. Strong secrecy asks that the total mutual information I(M:E) goes to 0,
3. Semantic security asks that the total mutual informaiton I(M:E) goes to 0 for any distribution of the message M (and thus in particular for all distributions that pick any of two chosen messages with 1/2 probabilty)

Then there are the almost-equivalent respective indistiguishablity types  of security requirements (below |P-Q|_1 is the statistical distance and Exp_M is expectation value over M)

1. average indistinguishability 1/n Exp_M | P_{E|M} - P_E |_1 for a uniform message M goes to 0 (again average refers over the blocklegth n, clearly there is also the average over M)
2. total indistiguishability Exp_M | P_{E|M} - P_E |_1 for a uniform message M goes to 0
3. indistinguishability |P_{E|m} - P_{E|m'}|_1 for any two messages m and m' goes to 0.

Each of the indistiguishabilities can also be written using KL digvergence instead of statistical distance, in which case the conditions are exactly equivalent to their mutual information versions.

Strong secrecy is the standard security requirement considered in information-theoretic security, while semantic security is the minimum requirement considered in computational security.
Information-theoretic (physical-layer) security differs from computational security in that the secrecy is guaranteed irrespective of the power of the adversary, while in computational security E is computationally bounded. Computational security also assumes that the message is at least of a certain length for the schemes to work, and thus if the message to be secured is too small it needs to be padded to a larger message.

In practice, information theoretic security is expensive, because the messages that can be secured can be only as long as the keys that can be generated. However, in identification only a very small part of the message needs to be secured, which in computational security triggers padding and thus waste, but on the other side makes information-theoretic security accessible and not so expensive.

At the same time, the security of identification implicitly requires semantic security. It has been known for a while that hash functions provide information-theoretic strong secrecy. However, because the standard for information-theoretic security has been strong secrecy, before https://arxiv.org/abs/2102.00983 no efficient functions where known to provide information-theoretic semantic security.
We need an implementation of these type of functions so that we can integrate information-theoretic security into our identification project.

This paper uses the partial transpose as a tool to derive upper bounds on device-independent QKD
https://arxiv.org/abs/2005.13511
In this project the goal is to try to generalize the above to the other tools like the reallignment criterion:
https://arxiv.org/abs/quant-ph/0205017
https://arxiv.org/abs/0802.2019

basics of quantum information/quantum formalism

Generalize semantic security of classical-quantum channels to infinite dimensional channel (not necessarily gaussian)

• [1] finite dimensional classical-quantum case
  https://arxiv.org/abs/2001.05719
• finite and infinite dimensional classical case
  https://arxiv.org/abs/1811.07798
• [this subpoint can be a project by itself] the finite dimesional case needs to be recast into smooth-max information (instead than Lemma 5.7 of [1]) as the classical case does, this paper proves properties of the smooth-max-inf in finite dimension that we would need for that
  https://arxiv.org/abs/2001.05719
• papers regarding the capacity for infinite dimensional channels
  http://arxiv.org/abs/quant-ph/9912067v1
  http://arxiv.org/abs/quant-ph/0408009v3
  http://arxiv.org/abs/quant-ph/0408176v1

quantum information theory

Asypmtotic continuity is a property in the form of inequalities (classically known also as inequalities of the reverse-Pinker type) that is necessary to prove upper bounds on operational capacities.

The (quantum) relative entropy (also known as quantum divergence and classically also known as Kullbackt-Leibler divergence), can be used to define various entanglment measures many of which have a proven asymptotic continuity.

Of particular interest are the restricted quantum relative entropies defined by Marco Piani (https://arxiv.org/abs/0904.2705), many of which satisfy asymptotic continuity (A.S.)

• https://arxiv.org/abs/quant-ph/9910002
• https://arxiv.org/abs/quant-ph/0203107
• https://arxiv.org/abs/quant-ph/0507126
• https://arxiv.org/abs/1210.3181
• https://arxiv.org/abs/1507.07775
• https://arxiv.org/abs/1512.09047

In the above there are maybe 2-3 different proof styles.
We can group the results in the above as follows:

• A.S. for entropy, conditional entropies, mutual information, conditional mutual information
• A.S. for relative entropies with infimum over states on the second argument
• A.S. relative entropies with infimum over state *and maximization over measurement channels*

The goal of the project is to generalize the last case to asymptotic continuity for relative entropies with infimum over state and maximization over *general* channels.

• Partial results toward this goal can be found in the appendix of my PhD thesis: http://web.math.ku.dk/noter/filer/phd18rf.pdf
• Such a result would have immediate applications to this paper: https://arxiv.org/abs/1801.02861

Possible new proof directions are

• using Renyi α-realtive entropies with the limit α->1
• using Kim's operator inequality from
  https://arxiv.org/abs/1210.5190
  to get an operator inequality looking like a reverse strong subadditivity (see https://www.youtube.com/watch?v=P3-xI1u1Y2s for a good overview and in particular at minute 31:20 for the reverse SSA)

Knowledge of quantum information is highly recommended/required.
Knowledge of matrix analysis will be a strong advantage.

relevant papers
https://arxiv.org/abs/1310.6043
https://arxiv.org/abs/1304.5944
https://arxiv.org/abs/1310.6045
https://arxiv.org/abs/1703.05876
https://arxiv.org/abs/1303.6357
background papers
https://ieeexplore.ieee.org/document/7509657

Knowledge of quantum theory as provided by the course Algorithms in Quantum Theory or similar

The goal of this work is to try to upper bound the device-independent distillable key in terms of locally restricted relative entropy of entanglement (an entanglement measure).

The following are relevant works/articles

• works toward even *a definition* of device independent distillable key
  https://arxiv.org/abs/2005.13511
  https://arxiv.org/abs/2005.12325
  https://arxiv.org/abs/1810.05627
• works relating distillable entanglement and distillable key to locally restricted relative entropy measures
  https://arxiv.org/abs/1609.04696
  https://arxiv.org/abs/1402.5927
• the first definition of restricted relative entropies
  https://arxiv.org/abs/0904.2705
• important properties of restricted relative entropies, and some overview of entanglement measures
  https://arxiv.org/abs/1210.3181
• my PhD thesis
  http://web.math.ku.dk/noter/filer/phd18rf.pdf

Strong background in quantum theory is required, preferably in quantum information theory, which is not covered by the course Algorithms in Quantum Theory

In this project, we investigate the interplay between redundancy and straggler tolerance in distributed learning.

The setting is that of a main node distributing computational tasks to available workers as part of a machine learning algorithm, e.g., training a neural network. Waiting for all workers to return their computations suffers from the presence of stragglers, i.e., slow or unresponsive nodes. Mitigating the effect of the stragglers can be done through the use of redundancy or by leveraging the properties of the convergence of the machine learning algorithm. 

The goal of this work is to compare when redundancy is needed. In this case, we aim to first analyze the convergence speed with and without redundancy. Then, we aim at design schemes that adaptively increase the redundancy to speed up the convergence.

Further reading: 

R. Bitar, M. Wootters and S. El Rouayheb, Stochastic Gradient Coding for Straggler Mitigation in Distributed Learning, IEEE Journal on Selected Areas in Information Theory (JSAIT), Vol. 1, No. 1, May 2020. arXiv:1905.05383

S. Kas Hanna, R. Bitar, P. Parag, V. Dasari and S. El Rouayheb, Adaptive Stochastic Gradient Descent for Fast and Communication-Efficient Distributed Learning, preprint, arXiv:2208.03134.

Knowledge in the following topics:

• Probability Theory
• Gradient descent and stochastic gradient descent
• Coding theory

Independence and motivation to work on a research topic

Knowledge of implementing neural networks is a plus

Space-division multiplexing (SDM), which consists in exploiting multimode fibers (MMFs) or multicore fibers (MCFs) instead of single mode ones, is one of the future optical communications architectures to increase data rates and network planning flexibility. The nonlinear properties of MCFs are of primary interest in assessing the usefulness of SDM against the current network. With this thesis, the student has the chance to work on a state-of-the-art topic in the field of optical communication systems, and progress quickly thanks to a tight (if desired) supervision. Would you be curious to know more about it? If so, just get in touch with me at paolo.carniello@tum.de (personal page https://www.ce.cit.tum.de/lnt/mitarbeiter/doktoranden/carniello/).

-some knowledge on optical communications systems (e.g., Optical Communication Systems or Simulation of Optical Communication Systems Lab)

-some knowledge about communications engineering topics

Space-division multiplexing (SDM), which consists in exploiting multimode (MMF) or multicore fibers instead of single mode ones, is one of the future architectures to increase data rates and network planning flexibility. A desired working condition for SDM is the so called strong-coupling linear regime, which is however not intrinsically achievable in common MMFs. With this topic, the student has the chance to investigate if it would be achievable with some new design. If you are curious about it, just send a mail to paolo.carniello@tum.de.

Basics of Optical Communication Systems (see https://www.ce.cit.tum.de/en/lnt/teaching/lectures/optical-communication-systems/)

During the internship, the student will be researching the application of Neural Network-based signal predistortion to mitigate the effects of fiber chromatic dispersion in direct detection systems.

• basic Python skills beneficial

Novel use cases for mobile communication networks include the aggregation of large amounts of data, which is stored in a distributed manner across network users. For instance, Federated Learning requires the aggregation of machine learning model updates from contributing users.

Over-the-Air (OtA) computation is an approach with the potential to drastically reduce the communication overhead of wireless distributed data-processing systems (e.g. Federated Learning). It exploits the multiple-access property and linearity of the wireless channel to compute sums of pre-processed data by the channel. This important property at the same time opens great opportunities for eavesdroppers to learn about the transmitted signal. If the legitimate receiver shall have exclusive access to the computation result, it is crucial to employ additional security measures.

Artificial noise can be employed to secure the communication. This noise is either generated by dedicated users jamming the communication [3], or by jointly designing the noise contribution of each user, [1][2]. The latter approach makes it possible to minimize the distortion at the legitimate receiver, but requires a centrally coordinated noise design. Therefore, an open problem is how to allow for the distributed design of artifical noise.

[1] Maßny, Luis, and Antonia Wachter-Zeh. "Secure Over-the-Air Computation using Zero-Forced Artificial Noise." arXiv preprint arXiv:2212.04288 (2022).
[2] Liao, Jialing, Zheng Chen, and Erik G. Larsson. "Over-the-Air Federated Learning with Privacy Protection via Correlated Additive Perturbations." arXiv preprint arXiv:2210.02235 (2022).
[3] Yan, Na, et al. "Toward Secure and Private Over-the-Air Federated Learning." arXiv preprint arXiv:2210.07669 (2022).

- basic knowledge in statistics and estimation theory
- basic knowledge about linear wireless channels

Identification is a communication scheme that allows rate doubly exponential in the blocklemght, with the tradeoff that identities cannot be decoded (as messages do) but can only be verified.

• https://ieeexplore.ieee.org/document/42172
• https://ieeexplore.ieee.org/document/42173

Identification codes can be achieved by first removing the errors from the channel with regular transmission channel coding, and then sending a challenge though the corrected channel. For every identity i, The channenge is generated by picking a random input m and computing the corresponding output T_i(m) using a function T_i that depends on the identity. The challenge is then the pair m,T_i(m) and the receiver wanting to verify an identity j will verify whether j=i by testing the challenge. This is done by recomputing the output with T_j and verifying whether T_j(m)= T_i(m). The errors are reduced by ensuring that the various functions collide on a small fraction of the possible inputs.

It turns out that choosing good sets of funtions {T_i} is the same as choosing error-correction codes {c_i} with large distance, where now each codeword c_i defines a function by mapping positions m (sometimes called code locators) to symbols c_im of the codeword.
We can thus construct identification codes by choosing error-correction codes where we are only interested in the performance of the error correction encoders (we are not interested in the error-correction decoder or error-correction codes).

From previous work we have a fairly efficient implementation based Reed-Muller code which can be found at

• https://arxiv.org/abs/2107.07649
• https://arxiv.org/abs/2107.06801
• https://arxiv.org/abs/2007.06372

Secrecy in this identification codes has also been implemented in unpublished work. Furthermore, the theoretical work on Identification with PUF's has been done in

• https://ieeexplore.ieee.org/document/8445910
• https://ieeexplore.ieee.org/document/8600405

The goal of the project will be to bridge the three topics and create practical and efficient secret identification codes in the PUF setting.

The working language will be in English.

Environment: this is a project in collaboration with LTI. At LNT and LTI there is currently a lot of funding for research in identification. Therefore you will find a large group of people that might be available for discussion and collaboration.

We consider the problem of running a distributed machine learning algorithm on the cloud. This imposes several challenges. In particular, cloud instances may have different performances/speeds. To fully leverage the performance of the instances, we want to characterize their speed and potentially use the fastest ones. To explore the speed of the instances while exploiting them (assigning computational tasks), we use the theory of multi-armed bandits (MABs).

The goal of the research intership is to start by implementing existing theoretical algorithms [1] and possibly adapting them based on the experimental observations.

[1] M. Egger, R. Bitar, A. Wachter-Zeh and D. Gündüz, Efficient Distributed Machine Learning via Combinatorial Multi-Armed Bandits, submitted to IEEE Journal on Selected Areas in Communications (JSAC), 2022.

• Information Theory
• Machine Learning Basics
• Python (Intermediate Level)

As part of an ongoing project with Huawei we are looking into quantum machine learning algorithms applied to decoding at the end of an optical fiber in the non-linear regime.

So far we have tried only the quantum version of k-mean clustering, however the goal is to test further quantum algorithms, in particular quantum support vector machines next, and their classical quantum-inspired counterpart.

The projects will involve reading the literature on quantum machine learning algorithms and quantum-inspired algorithms, find or come up with an implementation (this will involve the use of quantum libraries, in particular so far we have use qiskit), and benchmark the performance.

Knowledge of quantum mechanics or quantum information is highly recommended.

Identification is a communication scheme that allows rate doubly exponential in the blocklemght, with the tradeoff that identities cannot be decoded (as messages do) but can only be verified.

• https://ieeexplore.ieee.org/document/42172
• https://ieeexplore.ieee.org/document/42173

The double exponential growth presents various challenges in the finite regime: there are heavy computational costs introduced at the encoder and decoder and heavy trade-offs between the error and the codes sizes.

The ultimate goal is to find a fast, reliable implementation while still achieving large code sizes.

Identification codes can be achieved by first removing the errors from the channel with regular transmission channel coding, and then sending a challenge though the corrected channel. For every identity i, The channenge is generated by picking a random input m and computing the corresponding output T_i(m) using a function T_i that depends on the identity. The challenge is then the pair m,T_i(m) and the receiver wanting to verify an identity j will verify whether j=i by testing the challenge. This is done by recomputing the output with T_j and verifying whether T_j(m)= T_i(m). The errors are reduced by ensuring that the various functions collide on a small fraction of the possible inputs.

It turns out that choosing good sets of funtions {T_i} is the same as choosing error-correction codes {c_i} with large distance, where now each codeword c_i defines a function by mapping positions m (sometimes called code locators) to symbols c_im of the codeword.
We can thus construct identification codes by choosing error-correction codes where we are only interested in the performance of the error correction encoders (we are not interested in the error-correction decoder or error-correction codes).

Your task will be implementing the identification codes described in

• https://ieeexplore.ieee.org/abstract/document/782144

aiming at the fastest implementation, and testing their performance in comparison to other current implementations.

For reference, our previous work on identification based on Reed-Solomon and Reed-Muller code can be found at

• https://arxiv.org/abs/2107.07649
• https://arxiv.org/abs/2107.06801
• https://arxiv.org/abs/2007.06372

The coding will be in Python/Sagemath.
The working language will be in English.

Environment: we collaborate with LTI. At LNT and LTI there is currently a lot of funding for research in identification. Therefore you will find a large group of people that might be available for discussion and collaboration.

Identification is a communication scheme that allows rate doubly exponential in the blocklemght, with the tradeoff that identities cannot be decoded (as messages do) but can only be verified.

• https://ieeexplore.ieee.org/document/42172
• https://ieeexplore.ieee.org/document/42173

The double exponential growth presents various challenges in the finite regime: there are heavy computational costs introduced at the encoder and decoder and heavy trade-offs between the error and the codes sizes.

The ultimate goal is to find a fast, reliable implementation while still achieving large code sizes.

Identification codes can be achieved by first removing the errors from the channel with regular transmission channel coding, and then sending a challenge though the corrected channel. For every identity i, The channenge is generated by picking a random input m and computing the corresponding output T_i(m) using a function T_i that depends on the identity. The challenge is then the pair m,T_i(m) and the receiver wanting to verify an identity j will verify whether j=i by testing the challenge. This is done by recomputing the output with T_j and verifying whether T_j(m)= T_i(m). The errors are reduced by ensuring that the various functions collide on a small fraction of the possible inputs.

It turns out that choosing good sets of funtions {T_i} is the same as choosing error-correction codes {c_i} with large distance, where now each codeword c_i defines a function by mapping positions m (sometimes called code locators) to symbols c_im of the codeword.
We can thus construct identification codes by choosing error-correction codes where we are only interested in the performance of the error correction encoders (we are not interested in the error-correction decoder or error-correction codes).

Your task will be speeding up the current implementations based on Reed-Solomon and Reed-Muller codes:

• https://arxiv.org/abs/2107.07649
• https://arxiv.org/abs/2107.06801
• https://arxiv.org/abs/2007.06372

The coding will be in Python/Sagemath.
This work can accomodate multiple students.
The working language will be in English.

Environment: we collaborate with LTI. At LNT and LTI there is currently a lot of funding for research in identification. Therefore you will find a large group of people that might be available for discussion and collaboration.

Nachrichtentechnik 2

This paper uses the partial transpose as a tool to derive upper bounds on device-independent QKD
https://arxiv.org/abs/2005.13511
In this project the goal is to try to generalize the above to the other tools like the reallignment criterion:
https://arxiv.org/abs/quant-ph/0205017
https://arxiv.org/abs/0802.2019

basics of quantum information/quantum formalism