I am a research scientist/postdoc at the Chair of Information Theory and Machine Learning at Technische Universität Dresden. My research studies information flow in neural and communication systems: which parts are represented, which parts are protected, and what can be recovered under structural, statistical, computational, or adversarial constraints.

This perspective grew out of work on communication systems, including interference networks [1], wiretap channels [2], mutual-information estimation [3], neural channel coding [4], and generative channel models [5]. The current focus is on learned systems, where information flows through representations, optimization dynamics, and learned stochastic mechanisms.

I did my Dr.-Ing. (PhD) at Technische Universität Berlin, advised by Gerhard Wunder. My thesis was about deterministic models for capacity approximations in interference networks and physical layer security. I received the M.Sc. degree in electrical engineering from Technische Universität Berlin in 2012 and the B.Sc. degree in electrical engineering from Hochschule Furtwangen University in 2010.

Email / Google Scholar / GitHub / LinkedIn / ORCID

Research

The research problem is information flow under constraints: how structure changes what can be transmitted, hidden, estimated, or recovered. In my earlier work, these constraints were physical, algebraic, or communication-theoretic: interference, secrecy requirements, unknown channels, coding structure, and limited computational resources. This line includes channel coding [3], wiretap coding [6], interference networks [1], and mutual-information estimation [7].

In learned systems, neural networks induce implicit information-processing mechanisms through representations, optimization paths, and model outputs. This raises questions about leakage and interpretability: what is represented, what is discarded, what leaks, and what can be recovered from limited observations?

Communication systems provide a precise foundation for studying learned information processing. In channel coding, information must survive noise. In wiretap coding, information must be hidden from an adversary. In neural channel coding, recoverable representations are learned under strict noise, latency, or compute constraints. Modern learned systems face related questions when channels arise implicitly from architectures, objectives, or data.

Research Themes

• Information-theoretic structure and approximation: Complex systems can often be replaced by structured approximations that preserve the relevant information flow. In communication networks, this appears in deterministic models, interference geometry, capacity approximations, and duality relations [1] [8]. In learned systems, the analogous question is which task-relevant, private, or hidden information survives abstraction and compression.
• Security and privacy: My work on wiretap coding and physical-layer security studied reliable communication under adversarial inference constraints [2] [6]. The same tools also apply to leakage in neural systems, where adversarial inference can exploit outputs, representations, gradients, or training protocols.
• Estimating information in learned systems: Mutual information anchors communication theory, but it becomes difficult to estimate in high-dimensional learned systems. I study neural mutual-information estimation, estimator behavior, and information diagnostics for learned representations [7] [9]. This line of work supports the broader question of how to measure what is preserved or leaked in implicit channels.
• Learned channels and neural communication: Neural communication systems provide a controlled setting for studying representation and recovery under noise and compute constraints. Recent work uses diffusion models as learned channel approximations and recurrent architectures as scalable neural coding mechanisms [4] [5] [10].

Selected Projects and Code

• Condition-Wise Sinkhorn Drifting: Condition-preserving one-shot channel simulation with Sinkhorn drifting over conditional output laws. A GLOBECOM version has been accepted. [project page] [paper]

• MinDiffusion: Diffusion models as learned stochastic channel approximations for end-to-end communication-system training. [slides]

• Wireless Encoding with MI Estimation: Neural mutual-information estimation as a tool for optimizing and diagnosing information flow in channel coding.

• Neural Gaussian Wiretap Channel: Neural coding under adversarial observation, connecting physical-layer security with hidden information flow.

Recent News

• August 2026. A GLOBECOM version of “Condition-Wise Sinkhorn Drifting for One-Shot Learned Channel Simulation” was accepted.
• February 2026. Our collaborative paper “AI/ML-Driven 6G Network Solutions with Energy Efficiency Considerations” appeared in IEEE Access.

• January 2026. Our collaborative paper “6G PHY: Insights From 6G-ANNA Research Initiative” appeared in IEEE Open Journal of the Communications Society.

• May 2025. My paper “MinGRU-Based Encoder for Turbo Autoencoder Frameworks” with Rafael Schaefer appeared at ICMLCN 2025.
• May 2024. My colleague Muah Kim gave a tutorial about diffusion models at ICMLCN 2024 based on our work.
• January 2023. Our paper “Learning End-to-End Channel Coding with Diffusion Models” with Muah Kim and Rafael Schaefer was accepted.

Contact

Email: rick.fritschek at tu-dresden.de, rickfritschek at gmail.com