ResearchI am always open for collaboration to explore something new or related. General topics of interest are
Pointwise Maximal Leakage and Disclosure PreventionHow can useful information be disclosed while preventing sensitive inferences? Pointwise maximal leakage (PML) quantifies, on a logarithmic scale, the largest multiplicative improvement in an adversary’s expected gain after observing a particular disclosure. We develop its operational foundations and investigate how meaningful inferential privacy can coexist with useful data release. Our disclosure-prevention framework distinguishes revealing low-entropy features from protecting high-entropy features, providing a principled approach to privacy guarantees and mechanisms adapted to the data distribution. Complementary work introduces pointwise maximal cost (PMC) to quantify leakage to adversaries seeking to minimise costs. PML and PMC give operational interpretations to upper and lower bounds on information density, respectively. We show that local differential privacy requires both bounded PML and bounded PMC, clarifying the relationships between these privacy measures and the adversarial objectives they address.
Correlated Data and the Risk of a False Sense of PrivacyStatistical dependencies can allow sensitive information about an individual to be inferred from data about others. Using PML, we show that, for suitably correlated databases, a mechanism satisfying pure differential privacy can disclose almost as much information about an individual entry as directly releasing it. Our work on Nash equilibrium computation demonstrates how this issue can lead to unintended disclosure of players’ private information in distributed decision-making. By incorporating prior knowledge and correlations, PML enables a more informative assessment of these risks. The results highlight how applying privacy mechanisms without examining their inferential implications can create a false sense of privacy.
Optimal Privacy Mechanisms with Known and Estimated DistributionsHow can we maximise the usefulness of disclosed data under a PML privacy constraint? We characterise optimal mechanisms for a broad class of convex utility functions, obtaining closed-form solutions in several settings and a linear-programming formulation for general cases. We then address the practical situation where the data-generating distribution must be estimated from samples. By accounting for uncertainty in this estimate, we obtain probabilistic privacy guarantees and tractable mechanism-design methods without requiring exact knowledge of the prior. Together, these results connect optimal privacy–utility trade-offs with data-driven implementation.
Privacy Mechanisms with Worst-Case Utility GuaranteesIn some applications, a privacy mechanism must avoid unacceptable outputs rather than merely provide good average performance. We design discrete PML mechanisms that maximise worst-case utility by assigning zero probability to undesirable input–output pairs while preserving the privacy constraint. We derive a computationally efficient utility-safe mechanism, establish its optimality under an output-support constraint, and characterise the general optimal solution. This enables mechanism design with explicit guarantees on both privacy leakage and the utility of every possible disclosure.
PML on General Alphabets and Markov ProcessesPrivacy assessment must accommodate continuous measurements, stochastic processes, and other objects beyond finite datasets. We extend PML to random variables on arbitrary probability spaces while preserving its operational interpretation through adversarial estimation and gain functions. Building on this foundation, we develop decomposition-based bounds for finite-alphabet Markovian systems, including open-loop controlled, hidden, and partially observable Markov processes. These bounds exploit conditional independence and incorporate an information-density correction for statistical dependencies, providing analytical tools for assessing privacy leakage from temporally correlated observations.
Bridging Law and TechnologyHow can legal principles inform the mathematical design of privacy-preserving systems? In collaboration with legal scholars, we explore connections between principles from law, mathematics, and statistics to bridge the gap between legal requirements and technical implementation. Our work on calculated privacy connects the data minimisation principle with sufficient statistics: retaining the information needed for a specified task while limiting unnecessary disclosure. These connections provide a basis for interdisciplinary discussion of privacy guarantees and risk assessment. Complementary work examines machine unlearning and the right to be forgotten, considering how technical notions of removing information relate to legal expectations of erasure.
Privacy-Preserving Wireless LocalizationWireless localization should provide useful positioning information without exposing unnecessarily precise locations. We investigate privacy protection within the network infrastructure through beamspace processing and local perturbation before information reaches a fusion centre. Using the Cramér–Rao bound, we design mechanisms that balance localization performance against prescribed lower bounds on the error of unbiased position estimators. This places privacy control within the localization architecture without requiring additional privacy processing on the user device.
Joint Communication, Control, and CoordinationHow do communication constraints shape cooperation between decision-makers? We use coordination coding to study systems in which actions both influence physical states and convey information. For vector-valued Witsenhausen setups with causal encoding and noncausal decoding, we characterise power–estimation trade-offs and develop coding strategies with asymptotically zero estimation cost at reduced power. Complementary work establishes first-order optimality of a low-power strategy. We also study coordination over input-driven Markov channels, deriving inner and outer bounds that explicitly account for channel memory. These results clarify the roles of information patterns, feedback, and temporal dependencies in joint communication and control.
Understanding and Modelling Sensor DriftReliable long-term sensing requires distinguishing environmental influences from changes within the sensor itself. We develop a stochastic framework that separates reversible environmental effects from instrumental drift in NDIR gas sensors. A constrained additive model enables identifiable decomposition and improves the interpretation and correction of measurement errors. Complementary work investigates mean-reverting instrumental drift using Ornstein–Uhlenbeck models, revealing sensor-specific behaviour that can inform calibration strategies.
Calibration Algorithms for Sensor Networks and Embedded DevicesReliable low-cost sensing requires calibration methods that exploit available measurements and remain feasible on limited hardware. At the network level, we develop adaptive similarity search to identify periods when neighbouring sensors observe the same phenomenon, improving calibration compared with selection based on Pearson correlation. Within individual devices, we decompose inverse sensor transfer functions into components that can be approximated separately. This enables systematic trade-offs between compensation accuracy, memory, and computation, linking network-assisted calibration with efficient embedded implementation.
Strong Coordination in Distributed NetworksHow much communication and shared randomness are needed for distributed agents to generate actions with a prescribed joint distribution? We study strong coordination, requiring the distribution of entire sequences to approach a target distribution in total variation. For a discrete memoryless channel with causal encoding and noncausal decoding, we establish an exact characterisation of the achievable common-randomness rates and joint distributions. In multi-terminal networks, we derive achievable regions and matching bounds in special cases, demonstrate benefits from encoder cooperation, and incorporate secrecy against external eavesdroppers. These results clarify how information patterns, communication, and shared randomness determine the possibilities for coordinated behaviour.
Privacy-preserving LearningDue to privacy reasons, disclosure of data might not be possible so that value of the data is not exploited. It is therefore important to develop and analyze data analysis tools with disclosure control and provable privacy guarantees for learning and sharing of data. The goal is to develop clever mechanisms that maximize the utility while satisfying the suitable privacy constraints.
Biometric Identification SystemsBiometric data brings convenience in systems, but it also invokes privacy and secrecy issues. Physical layer security provides strong strong privacy and security guarantees. Thus, how to design optimal biometric identification systems and what are the optimal trade-offs?
Data-driven Calibration AlgorithmsThe basic working mechanism of a sensor is to convert a specific physical quantity to an electrical signal. The physical model of the sensor however might not incorporate all dependencies of environmental factors. One vital approach is adjust a calibration parameter to compensate for the remaining environmental factors. How to design efficient algorithms to learn the dependency and provide mechanisms that provide a belief function on the sensor measurements?
Fundamental Bounds in Distributed Decision-makingIn 1968, Witsenhausen formulated a two-stage control problem for which he showed that non-linear policies outperform the best linear policy. Therewith, he disproved the conjecture that the optimality of linear strategies for centralized linear quadratic Gaussian control problems extends to decentralized systems. The optimal solution of this toy problem remains unsolved, but it is an interesting study object since it points on a very fundamental issue. We take a fresh approach approach deriving fundamental bounds using coordination coding results.
Smart Meter PrivacyEnergy consumption profiles contain lots of private information about consumer habits and actions. Thus, privacy becomes more and more an issue with increasing resolution of smart meters in future smart grids. How can we assess the privacy risk and how can we protect the consumer privacy? Can we manipulate the energy consumption profile to enhance the privacy level? What are the fundamental limits to protect against un-authorised inference on consumers behaviour and habits?
IdentificationIdentification of a user or item is a key task of a search in a database, e.g. for authentication. However with high dimensional data, the storage size and search complexity become the critical parameters. If the identification additionally deals with personal data, then we also face a privacy problem. To deal with such problems, we study fundamental properties of a two stage identification system with pre-processing.
Secure EstimationEstimation and tracking the state of a dynamical system is essential for future cyber-physical systems. However, we may want to keep the state keep secret. How can we achieve a secure communication over an uncertain channel so that an unauthorized receiver will not be able to track the state? What are the fundamental necessary and sufficient conditions?
Distributed Detection with PrivacyConsider a distributed detection problem where links are attacked by an eavesdropper. How to achieve a privacy-aware system design considering a detection-theoretic operational privacy metric? What kind of detection strategies are sufficient to consider for optimization? Some selected publications:
Optimal Transmit Strategies with Per-Antenna Power ConstraintsThe capacity of wireless channels can be significantly increased if we use multiple antennas at the transmitter and receiver. Since each transmit antenna has its own power amplifier, it is reasonable to assume that we have per-antenna power constraint. Thus, what are the optimal transmit strategies taking this into account? How can we compute the optimal transmit strategies?
Secure Source Coding ProblemsConsider a distributed source coding setup where the communication is eavesdropped. What is the optimal coding strategy limiting the information leakage to an eavesdropper? What are bounds on the optimal trade-off between rates over public and private channels, distortion, and information leakage? Some selected publications:
Broadcast Channel with RX Message CognitionConsider communication problems over a broadcast channel where the receivers have (partial) knowledge about messages to be transmitted to other receivers. What is the optimal coding strategy? Considering (vector-valued) Gaussian channels, what is the optimal transmit strategy/input distribution? Some selected publications:
Multi-terminal Networked ControlConsider a multi-terminal Gaussian closed-loop control setup where the plant observations are communicated over Gaussian channels to the controller(s). What are necessary and sufficient conditions for mean-square stability? What are suitable/optimal communication and control strategies?
Approximation of Gaussian Mixture DistributionConsider a Gaussian mixture distribution with a very large number of components. How can I approximate the entropy of such distribution? Which terms are most relevant and how to merge and prune other terms?
Complexity efficient receiver algorithm for IEEE802.11pWireless communication in vehicular environment is challenging due to the variation of the channel state. How can we improve the receiver performance in such environments? How can we efficiently estimate the channel state and decode the payload in IEEE802.11p communication?
|