Mitarbeiter Lehrstuhl/Einrichtung der Fakultät für Informatik und Mathematik
Refine
Year of publication
Document Type
- Doctoral Thesis (57)
- Preprint (11)
- Article (1)
- Conference Proceeding (1)
Has Fulltext
- yes (70)
Keywords
- Maßtheorie (12)
- Graphenzeichnen (8)
- Iteriertes Funktionensystem (6)
- Quantisierung (6)
- Graph (5)
- Modellierung (5)
- Programmanalyse (5)
- Fraktale Dimension (4)
- Hausdorff-Dimension (4)
- Hausdorff-Maß (4)
In this thesis a new approach to building product recommender systems is introduced. By using a customer-centric dialogue, the customers' preferences are elicited. These are the basis for inferring utility estimations about the desired technical properties of the products in question. Systems built this way can both operate autonomously, e.g., in an online store, and support a salesperson directly at the point-of-sale. The core of the approach is formed by a layered domain description that models customer stereotypes and needs, product attributes, the products themselves, and the causal interrelations between customer and product properties. Maintenance of the domain description, i.e., keeping the model up-to-date in face of frequent changes, is facilitated by the clear separation of concerns provided by the layered structure. In fact, the most frequently used class of updates can be handled in an entirely automated way if some constraints are satisfied. On a high level of abstraction, the system behavior is described by State Charts that are parameterized according to the domain description. Those parts of the system description where State Charts would be too imprecise are implemented by separate components realizing the required complex semantics. From the domain description, a Bayesian network is generated that forms the core of the inference engine of the recommender system. The network essentially controls the system-initiated dialogue flow and the recommendation process. Due to the characteristics of Bayesian networks, it is possible to respond to user-initiated dialogue steps in a natural way. Moreover, an explanation of the current recommendation can be generated without having to explicitly encode additional information in the modeling layer. Finally, a database structure and the SQL queries necessary to obtain recommendations can be inferred from the corresponding parts of the domain description. Instantiation of the system to a specific business domain is supported by a dedicated maintenance application that hides the complexities of the underlying algorithms. Thus, day-to-day system updates by non-technical domain experts, e.g., product managers, are facilitated. The developed concepts were implemented in cooperation with a local industry partner who intends to apply the recommender system in the field of mobile communications.
Up to a few years ago, the typical operation of a distributed architecture was modelled as the enactment of a collaborative protocol by networked nodes. In this context, all nodes were under the system designer’s control, faithfully executing the programmed behaviour. However, today’s networks are often characterized by a free aggregation of nodes. Thus, the possibility increases that a selfish party operates a node, which may violate the collaborative protocol in order to increase a personal benefit. If such violations differ from the system goals they can even be considered as attack. Current fault-tolerance techniques may weaken the harmful impact to some degree but they cannot necessarily prevent them. Furthermore, the several architectures differ in their fault-tolerance capabilities. This emphasizes the need for a systematic approach to achieve collaboration in distributed systems. In this PhD thesis we consider the problem of attaining a targeted level of collaboration in a distributed architecture deployed over rational selfish-driven nodes, which have interest in deviating from the communication protocol to increase a personal benefit. In order to reach this goal and to cover a broad spectrum of systems, we do not modify the architecture or communication protocol itself. Instead, we add a monitoring logic to inspect a node’s behaviour in terms of the correct interaction with the system. With this approach, the system designer needs to contrast several aspects such as the specific environmental circumstances, the inspection effort or the node’s individual preferences. Furthermore, he should consider the fact that each agent could be aware of the other agents’ preferences and selfishness, and perform strategic choices consequently. The natural frame for modelling such complex, interdependent and possibly interactive decision landscape is Game Theory (GT). In this context, the monitoring setup proposed in this thesis corresponds to a class of GT models known as Inspection Games (IG). Such games were introduced 1962 in their simplest formulation by Dresher in the context of non-proliferation treatises and arm control. They model the general situation where one inspector verifies through inspections the correct behaviour of another party, called inspectee. However, inspections are costly and the inspector’s resources are limited. Hence, a complete surveillance is not possible and an inspector will try to minimize the inspections. Finally, a game strategy combination (violating/inspecting or not) that is considered optimal by the parties represents a Nash equilibrium for the game. In this thesis, the initial IG model is enriched by the possibility of false negatives, i.e. the probability that a violation is not detected during an inspection. Both the initial and the enriched model remain abstract and can thus easily find interdisciplinary application. However, as solution approach in this thesis considering the context of distributed systems, it models the network participants’ strategy choice. As outcome, the IG model enables to calculate system parameters in order to shift the Nash equilibrium to the desired target collaboration. The approach is designed as framework. It can be therefore applied to any architecture considering, any selfish goal and any reliability technique. For sake of concreteness, we will discuss the IG approach by means of the illustrative case of a Publish/Subscribe (pub/sub) architecture. In this way messages over the communication infrastructure will have a specific associated semantics. The Inspection Game approach of this thesis secures the whole collaborative protocol in order to attain a correctly working system up to a specific degree (in the sense of collaboration). This represents a completely new way in terms of reliability mechanisms. Hence, this thesis can be considered as fundamental research. In order to enable a broad application, the generality of this approach is supported by further contributions. This is among others the software library RCourse for practical robustness evaluations of overlay networks and a simulation environment for further research of the abstract IG model. All developments will finally be published as open source software.
Program slicing is a technique to identify statements that may influence the computations in other statements. Despite the ongoing research of almost 25 years, program slicing still has problems that prevent a widespread use: Sometimes, slices are too big to understand and too expensive and complicated to be computed for real-life programs. This thesis presents solutions to these problems: It contains various approaches which help the user to understand a slice more easily by making it more focused on the user's problem. All of these approaches have been implemented in the VALSOFT system and thorough evaluations of the proposed algorithms are presented. The underlying data structures used for slicing are program dependence graphs. They can also be used for different purposes: A new approach to clone detection based on identifying similar subgraphs in program dependence graphs is presented; it is able to detect modified clones better than other tools. In the theoretical part, this thesis presents a high-precision approach to slice concurrent procedural programs despite that optimal slicing is known to be undecidable. It is the first approach to slice concurrent programs that does not rely on inlining of called procedures.
One of the most important algorithms for real quantifier elimination is the quantifier elimination by virtual substitution introduced by Weispfenning in 1988. In this thesis we present numerous algorithmic approaches for optimizing this quantifier elimination algorithm. Optimization goals are the actual running time of the implementation of the algorithm and the size of the output formula. Strategies for obtaining these goals include simplification of first-order formulas,reduction of the size of the computed elimination set, and condensing a new replacement for the virtual substitution. Local quantifier elimination computes formulas that are equivalent to the input formula only nearby a given point. We can make use of this restriction for further optimizing the quantifier elimination by virtual substitution. Finally we discuss how to solve a large class of scheduling problems by real quantifier elimination. To optimize our algorithm for solving scheduling problems we make use of the special form of the input formula and of additional information given by the description of the scheduling problem
Modern Web technology makes the dream of fully interactive and enriched video come true. Nowadays it is possible to organize videos in a non-linear way playing in a sequence unknown in advance. Furthermore, additional information can be added to the video, ranging from short descriptions to animated images and further videos. This affords an easy and efficient to use authoring tool which is capable of the management of the single media objects, as well as a clear arrangement of the links between the parts. Tools of this kind can be found rarely and do mostly not provide the full range of needed functions. While providing an interactive experience to the viewer in the Web player, parallel plot sequences and additional information lead to an increased download volume. This may cause pauses during playback while elements have to be downloaded which are displayed with the video. A good quality of experience for these videos with small waiting times and a playback without interruptions is desired. This work presents the SIVA Suite to create the previously described annotated interactive non-linear videos. We propose a video model for interactivity, non-linearity, and annotations, which is implemented in an XML format, an authoring tool, and a player. Video is the main medium, whereby different scenes are linked to a scene graph. Time controlled additional content called annotations, like text, images, audio files, or videos, is added to the scenes. The user is able to navigate in the scene graph by selecting a button at a button panel. Furthermore, other navigational elements like a table of contents or a keyword search are provided. Besides the SIVA Suite, this thesis presents algorithms and strategies for download and cache management to provide a good quality of experience while watching the annotated interactive non-linear videos. Therefor, we implemented a standard-independent player framework. Integrated into a simulation environment, the framework allows to evaluate algorithms and strategies for the calculation of start-up times, and the selection of elements to pre-fetch into and delete from the cache. Their interaction during the playback of non-linear video contents can be analyzed. The algorithms and strategies can be used to minimize interruptions in the video flow after user interactions. Our extensive evaluation showed that our techniques result in faster start-up times and lesser interruptions in the video flow than those of other players. Knowledge of the structure of an interactive non-linear video can be used to minimize the start-up time at the beginning of a video while minimizing an increase in the overall download volume.
In his famous paper Gersho stressed that the codecells of optimal quantizers asymptotically make an equal contribution to the distortion of the quantizer. Motivated by this fact, we investigate in this paper quantizers in the scalar case, where each codecell contributes with exactly the same portion to the quantization error. We show that such quantizers of Gersho type - or Gersho quantizers for short - exist for non-atomic scalar distributions. As a main result we prove that Gersho quantizers are asymptotically optimal.
For homogeneous one-dimensional Cantor sets, which are not necessarily self-similar, we show under some restrictions that the Euler exponent equals the quantization dimension of the uniform distribution on these Cantor sets. Moreover for a special sub-class of these sets we present a linkage between the Hausdorff and the Packing measure of these sets and the high-rate asymptotics of the quantization error.
In this dissertation we generalise the notion of level planar graphs in two directions: track planarity and radial planarity. Our main results are linear time algorithms both for the planarity test and for the computation of an embedding, and thus a drawing. Our algorithms use and generalise PQ-trees, which are a data structure for efficient planarity tests.
The majority of all security problems in today's Web applications is caused by string-based code injection, with Cross-site Scripting (XSS)being the dominant representative of this vulnerability class. This thesis discusses XSS and suggests defense mechanisms. We do so in three stages: First, we conduct a thorough analysis of JavaScript's capabilities and explain how these capabilities are utilized in XSS attacks. We subsequently design a systematic, hierarchical classification of XSS payloads. In addition, we present a comprehensive survey of publicly documented XSS payloads which is structured according to our proposed classification scheme. Secondly, we explore defensive mechanisms which dynamically prevent the execution of some payload types without eliminating the actual vulnerability. More specifically, we discuss the design and implementation of countermeasures against the XSS payloads Session Hijacking'', Cross-site Request Forgery'', and attacks that target intranet resources. We build upon this and introduce a general methodology for developing such countermeasures: We determine a necessary set of basic capabilities an adversary needs for successfully executing an attack through an analysis of the targeted payload type. The resulting countermeasure relies on revoking one of these capabilities, which in turn renders the payload infeasible. Finally, we present two language-based approaches that prevent XSS and related vulnerabilities: We identify the implicit mixing of data and code during string-based syntax assembly as the root cause of string-based code injection attacks. Consequently, we explore data/code separation in web applications. For this purpose, we propose a novel methodology for token-level data/code partitioning of a computer language's syntactical elements. This forms the basis for our two distinct techniques: For one, we present an approach to detect data/code confusion on run-time and demonstrate how this can be used for attack prevention. Furthermore, we show how vulnerabilities can be avoided through altering the underlying programming language. We introduce a dedicated datatype for syntax assembly instead of using string datatypes themselves for this purpose. We develop a formal, type-theoretical model of the proposed datatype and proof that it provides reliable separation between data and code hence, preventing code injection vulnerabilities. We verify our approach's applicability utilizing a practical implementation for the J2EE application server.
This thesis investigates, how placement variations of electronic devices influence the possibility of using sensors integrated in those devices for context recognition. The vast majority of context recognition research assumes well defined, fixed sen- sor locations. Although this might be acceptable for some application domains (e.g. in an industrial setting), users, in general, will have a hard time coping with these limitations. If one needs to remember to carry dedicated sensors and to adjust their orientation from time to time, the activity recognition system is more distracting than helpful. How can we deal with device location and orientation changes to make context sensing mainstream? This thesis presents a systematic evaluation of device placement effects in context recognition. We first deal with detecting if a device is carried on the body or placed somewhere in the environ- ment. If the device is placed on the body, it is useful to know on which body part. We also address how to deal with sensors changing their position and their orientation during use. For each of these topics some highlights are given in the following. Regarding environmental placement, we introduce an active sampling ap- proach to infer symbolic object location. This approach requires only simple sensors (acceleration, sound) and no infrastructure setup. The method works for specific placements such as "on the couch", "in the desk drawer" as well as for general location classes, such as "closed wood compartment" or "open iron sur- face". In the experimental evaluation we reach a recognition accuracy of 90% and above over a total of over 1200 measurements from 35 specific locations (taken from 3 different rooms) and 12 abstract location classes. To derive the coarse device placement on the body, we present a method solely based on rotation and acceleration signals from the device. It works independent of the device orientation. The on-body placement recognition rate is around 80% over 4 min. of unconstrained motion data for the worst scenario and up to 90% over a 2 min. interval for the best scenario. We use over 30 hours of motion data for the analysis. Two special issues of device placement are orientation and displacement. This thesis proposes a set of heuristics that significantly increase the robustness of motion sensor-based activity recognition with respect to sen- sor displacement. We show how, within certain limits and with modest quality degradation, motion sensor-based activity recognition can be implemented in a displacement tolerant way. We evaluate our heuristics first on a set of synthetic lower arm motions which are well suited to illustrate the strengths and limits of our approach, then on an extended modes of locomotion problem (sensors on the upper leg) and finally on a set of exercises performed on various gym machines (sensors placed on the lower arm). In this example our heuristic raises the dis- placed recognition rate from 24% for a displaced accelerometer, which had 96% recognition when not displaced, to 82%.