平成27年度 情報科学研究科 博士学位論文発表梗概


Verification Methods for Security against Inference Attacks on XML and Relational Databases

CHITTAPHONE PHONHARATH (1261020)


Access control is a traditional mechanism for confidentially restricting accesses to a database made by a user by dividing the queries into authorized and unauthorized ones, and restricting the portion of the data that can be retrieved and updated by the user. However, a user can obtain the result of an unauthorized query by combining the results of authorized queries and the meaning of the queries. Such an attack is called an inference attack. Quantitative notions to measure the security or privacy against inference attacks are needed because the perfect security or privacy cannot be achieved when indirect leakage or inference exists. Security against inference attacks has been the main motivation in this research. This presentation presents two studies on verifying the security against inference attacks on XML and relational databases.

Firstly, I will present the study of a static analysis problem on schema-level k-secrecy, which is a metric for the security against inference attacks on XML databases. Also, the discussion of the decidability and computational complexity of this problem. Secondly, I will present the proposal of the query-based l-diversity, which is a new privacy notion obtained by combining k-secrecy and l-diversity in the relational database setting, and two approaches to the problem of verifying the query-based l-diversity. I will also show the effectiveness and scalability of two approaches based on experimental results. Lastly, I will end this presentation with the conclusion and future work.


Automated Grammatical Error Correction Using Statistical Machine Translation Techniques with Revision Log of Language Learning SNS

水本智也 (1261016)


Recently, natural language processing research has begun to pay attention to second language learning. However, it is not easy to acquire large-scale learners’ corpora which is important to a research for second language learner by natural language processing. We present an attempt to extract a large-scale second language learners’ corpus from the revision log of a language learning social network service. This corpus is easy to obtain in large-scale, covers a wide variety of topics and styles, and can be a great source of knowledge for both language learners and instructors. I also demonstrate that the extracted learners’ corpus of Japanese/English as a second language can be used as training data for learners’ error correction using a statistical machine translation approach. For Japanese error correction, we proposed character-based SMT approach to alleviate the problem of erroneous input from language learners. We evaluate different granularities of tokenization to alleviate the problem of word segmentation errors caused by erroneous input from language learners. Experimental results show that the character-based model outperforms the word-based model.

For English, I conduct experiments in error correction targeting all types errors using statistical machine translation technique and I analyze the strength and weakness of grammatical error correction using statistical machine translation. I also propose the grammatical error correction method using discriminative reranking with POS/syntactic features. I show the effectiveness of reranking for grammatical error correction.


安全運転支援のための単眼カメラによる運転者状態推定の研究

松尾 治夫 (1261014)


自動車に車載されるセンサ、情報処理能力、制御技術の高度化により, 衝突防止や車線逸脱防止等の運転支援技術や, 通信機能を有する車載情報機器による情報提示技術の普及しつつある. しかし,自動車は運転者からの運転操作やスイッチや音声入力を受けて動作しているにすぎず, 安全運転支援のための運転者と自動車のインタラクション実現には, 運転者状態に応じた運転支援や情報提示が必要となる. これに対し従来,運転者の生体信号,カメラ映像および運転操作を用いて, 運転者の居眠り警報や脇見警報,運転者の視線や注意の方向に情報を提示する装置が提案されている. 本研究では,運転者の撮影したカメラ映像と運転操作の信号を用い, 運転者の覚醒低下や視認行動を評価し,従来よりも不安全行動を早期検出することで, 運転者に行動の変容を促す提案に適応できること示す.

本論文の1章では,事故防止の観点から自動車の略史について述べ, 運転者状態を得ることの重要性を示す.次に,運転者状態推定の研究を概観し, 本研究の意義を明らかにする. 2章では,画像合成により作成した運転者の顔画像を学習データとして 顔検出器を生成することで従来よりも光環境に対し堅牢な顔検出技術を提案する. 3章では,運転者の顔検出結果を用い,運転者ごとの較正を自動化した 顔向きと目の向き,目閉じの検出方法を提案し, これらを用いた不注意の検出と,居眠り運転の検出の可能性について述べる. 4章では,3章で得た顔の特徴に加えて,副次行動と,運転操作を用いることで, 運転者の覚醒低下を従来よりも早期に推定する方法について述べる. 5章では,3章と4章で推定した運転者状態に地図情報を組み合わせ, 車両運転イベントの自動抽出と記録と運転者への提示方法と提案し, 運転者状態推定技術が運転者の安全運転評価に有用なことを示す. 最後に6章で,本論文を総括し今後の課題と展望を述べる.


Unsupervised Anomaly Detection in Massive Traffic using S-transform and Renyi Divergence

Sirikarn Pukkawanna (1161207)


As sophisticated attacks grow exponentially, preserving security with signature-based Network Intrusion Detection Systems (NIDS) may not sufficient because they cannot detect unknown and new attacks.

In the presentation, I will introduce two novel network anomaly detection methods: (1) S-transform-based and (2) Renyi divergence-based methods. Both methods do not require the pre-defined signatures of targets and are able to detect unknown and new malicious and disruptive traffic with high accuracy and low false positive rates.

(1) Briefly, the S-transform-based method uses S-transform to convert a traffic signal (e.g., packet rate) to a time-frequency domain. The method then detects unusual time-frequency behavior caused by anomalies in the time-frequency domain. In the presentation, I will explain the method in detail and show the performance evaluation results obtained from testing the method with simulated and real-world backbone traffic from different datasets. Furthermore, I will show the comparison results between the proposed S-transform-based method and a popular Wavelet transform-based anomaly detection method in terms of accuracy and false positive rates. At the end of this part, I will show how to improve the detection performance of the S-transform-based method using sketch technique. The performance evaluation of the improved method with real-world backbone traffic is also presented in the presentation.

(2) For the Renyi divergence-based method, in summary, the method observes the port pair distribution of traffic and detects anomalies based on the R\'enyi divergence of the port pair distributions. The port pair distribution is a new statistical feature firstly proposed by us. In this presentation, I will present the details of the feature and method. In addition, I will show the performance comparison results between the proposed feature and four widely-used traffic features, namely the distributions of source IP, destination IP, source port, and destination port. I will also show the results obtained from comparing the performance between the proposed method and a Kullback-Leibler (KL) divergence-based anomaly detection method.

Lastly, I will discuss the limitations of both proposed methods and present the recommendations for future work.


Designing projected user interfaces as assistive technology for the elderly


HYRY JAAKKO MARKUS (1261022)


Old age brings several physical and cognitive challenges for using new technology, which complicates utilization of modern ICT for caretakers and family support and for daily task assist. One factor hindering the use of ICT are user interfaces (UI) that require prior knowledge of use metaphors that many elderly cannot learn to master.

Research on developing assistive technology exists, such as phones for the elderly with UIs that require prior knowledge and use experience. Latest research has introduced Ambient Assisted Living (AAL) concepts for users homes, for example using guidance projected in the environment. However, only a few empirical studies have attempted to define what type of projection-based UIs would be intuitive for the elderly and system design processes that would help in developing such AAL have not been researched thoroughly.

This work presents three design iterations and their empirical evaluations of which a body of knowledge is produced for designing and developing AALs with projected AR UIs. The first iteration has a sentence building UI implemented for a wearable Projector-Camera (ProCam) system, which had limits in technical suitability for the elderly. The second iteration changed the use metaphor to a simple icon-based menu, and produces a requirements guideline for UIs in AAL. In the final iteration the wearable was replaced with a fixed ProCam, allowing the elderly to make menu selections effectively and supports sequential tasks with visual guides such as taking medicine. Suitability of the new UI was compared between computer literate and elderly users, of whom many have memory and motor skill limitations. The comparison showed that the two groups are similar, but the elderly need a slower and direct UI. Assistance for the sequential tasks was found feasible. This work produced a set of UI-related and technical factors that AAL designers have to take into account when developing suitable projector-based AR systems for the elderly with memory problems. In addition, suggestions on how to conduct UI testing sessions with this user group can reduce amount of work and improve correctness of iterative development processes.


情報科学研究科 副専攻長