## Dissertations - ALL

May 2016

Dissertation

#### Degree Name

Doctor of Philosophy (PhD)

#### Department

Electrical Engineering and Computer Science

Yingbin Liang

#### Keywords

degraded broadcast channel, distribution-free test, kernel mean embedding, nonparametric anomaly detection, secrecy capacity region, secure communication

Engineering

#### Abstract

Two major security challenges in information systems are detection of anomalous data patterns that reflect malicious intrusions into data storage systems and protection of data from malicious eavesdropping during data transmissions. The first problem typically involves design of statistical tests to identify data variations, and the second problem generally involves design of communication schemes to transmit data securely in the presence of malicious eavesdroppers. The main theme of this thesis is to exploit information theoretic and statistical tools to address the above two security issues in order to provide information theoretically provable security, i.e., anomaly detection with vanishing probability of error and guaranteed secure communication with vanishing leakage rate at eavesdroppers.

First, the anomaly detection problem is investigated, in which typical and anomalous patterns (i.e., distributions that generate data) are unknown \emph{a priori}. Two types of problems are investigated. The first problem considers detection of the existence of anomalous geometric structures over networks, and the second problem considers the detection of a set of anomalous data streams out of a large number of data streams. In both problems, anomalous data are assumed to be generated by a distribution $q$, which is different from a distribution $p$ generating typical samples. For both problems, kernel-based tests are proposed, which are based on maximum mean discrepancy (MMD) that measures the distance between mean embeddings of distributions into a reproducing kernel Hilbert space. These tests are nonparametric without exploiting the information about $p$ and $q$ and are universally applicable to arbitrary $p$ and $q$. Furthermore, these tests are shown to be statistically consistent under certain conditions on the parameters of the problems. These conditions are further shown to be necessary or nearly necessary, which implies that the MMD-based tests are order level optimal or nearly order level optimal. Numerical results are provided to demonstrate the performance of the proposed tests.

The secure communication problem is then investigated, for which the focus is on degraded broadcast channels. In such channels, one transmitter sends messages to multiple receivers, the channel quality of which can be ordered. Two specific models are studied. In the first model, layered decoding and layered secrecy are required, i.e., each receiver decodes one more message than the receiver with one level worse channel quality, and this message should be kept secure from all receivers with worse channel qualities. In the second model, secrecy only outside a bounded range is required, i.e., each message is required to be kept secure from the receiver with two-level worse channel quality. Communication schemes for both models are designed and the corresponding achievable rate regions (i.e., inner bounds on the capacity region) are characterized. Furthermore, outer bounds on the capacity region are developed, which match the inner bounds, and hence the secrecy capacity regions are established for both models.

Open Access

COinS