オンライン上の不正アクセス防止のために，音声の聞き取り問題を提示して，ユーザが人間か機械かを判別する音声型

を提案する。日本語単語の隣接する2モーラを入れ替える手法（前後交替）と一つのモーラの子音を削除する手法（子音削除）で非語を作り，単語と判別する課題を行った。その際，機械（音声認識ツール）と人間で判別実験を行い，機械受入率と人間拒否率が最小化するよう，単語親密度，モーラ数，交替位置，削除位置を調整した。その結果，システムの安全性を表すF値が従来手法よりも向上し，の回答時間も短縮することができた。

抄録全体を表示

We propose a Multi-model that is composed of one sentence with typo and multiple pictures. This utilizes human's three abilities (inferring the meaning of the sentence with typo, understanding the meaning of pictures, and linking the meaning of the sentence and corresponding picture), and it is based on a hypothesis, "the longer of the sentence with typo, the time could be longer for a machine to link the meaning of the sentence and corresponding picture (= the shorter of the time of linking, the higher of the probability of human)." We found three findings from our experiment: (1) Examinees were able to link the meaning of the sentence to corresponding picture even there were some pictures with similar composition, (2) Typoglycemia is likely to be appeared on a sentence constructed by short and familiar words, (3) The time for linking was not exponentially increased even the length of the sentence was getting long (= It was not getting difficult for users in this situation).

抄録全体を表示

In previous research, a consisted of one sentence with typoglycemia and multiple pictures was proposed. However, questions presented to research subjects were manually made. Hence, the questions should be created automatically for practical use. Authors have proposed an automatic question making process and developed a software of creating the questions automatically. In our experiment with 50 research subjects, correct answer rate and answer time were the same as the results that the previous study has shown. It means that out process and software have almost reached levels of the results of the previous study.

抄録全体を表示

We propose new

image generation systems by using generative adversarial network (GAN) techniques to strengthen against solvers. To verify whether a user is human, images are widely used on the web industry today. We introduce two different systems for generating images, namely, the distance GAN (D-GAN) and composite GAN (C-GAN). The D-GAN adds distance values to the original images to generate new ones, and the C-GAN generates a image by composing multiple source images. To evaluate the performance of the proposed schemes, we used the breaker software as solver. Then, we compared the resistance of the original source images and the generated images against the solver. The results show that the proposed schemes improve the resistance to the solver by over 67.1% and 89.8% depending on the system.

抄録全体を表示

Malicious attackers on the Internet use automated attack programs to disrupt the use of services via mass spamming, unnecessary bulletin boarding, and account creation. Completely automated public turing test to tell computers and humans apart (

) is used as a security solution to prevent such automated attacks. is a system that determines whether the user is a machine or a person by providing distorted letters, voices, and images that only humans can understand. However, new attack techniques such as optical character recognition (OCR) and deep neural networks (DNN) have been used to bypass . In this paper, we propose a method to generate images by using the fast-gradient sign method (FGSM), iterative FGSM (I-FGSM), and the DeepFool method. We used the image provided by python as the dataset and Tensorflow as the machine learning library. The experimental results show that the image generated via FGSM, I-FGSM, and DeepFool methods exhibits a 0% recognition rate with ε=0.15 for FGSM, a 0% recognition rate with α=0.1 with 50 iterations for I-FGSM, and a 45% recognition rate with 150 iterations for the DeepFool method.

抄録全体を表示

is a kind of challenge response test, which is used to distinguish human users from malicious computer program such as bots. However, the attack technique called relay attack as a method to avoid the has been devised. This attack relays the challenges to remote human-solvers, let them to decode challenges. We used delay time that is caused by communications needed in relay attack. Our new uses this delay time between communications to prevent relay attacks. We constructed an experimental environment in which relay attack can be simulated, made a series of experiments in order to evaluate the performance of the proposed method.

抄録全体を表示

We propose a new

scheme that uses random dot patterns (RDPs) to prevent highly-developed bots attacks. Human beings can recognize a moving figure filled by a RDP from a background that is filled by another RDP; however, it is impossible to find such figures when they are stationary. Since image recognition by bots is usually carried out frame by frame, it is hard for bots to recognize such moving figures. The proposed scheme exploits this characteristic. Several experiments were carried out to confirm that the proposed scheme is usable enough and has enough resistance against bot attacks.

抄録全体を表示

The (Completely Automated Public Turing test to tell Computers and Humans Apart) idea is widely used as a HIP (Human Interactive Proof) for distinguishing between humans and computer programs. These are automated tests that humans can pass, but that current computer programs can't handle. Breaking a generally involves solving a difficult Artificial Intelligence problem. There are demands for new technologies that are stronger against automatic attacks by machines, without making it too hard for humans to pass the tests. In this paper, we propose a concept called the Oblivious , as a fifth-factor technology for a practical . The Oblivious uses tasks such as identifying the English alphabetic characters in a set of mixed alphabetic and non-alphabetic characters, or counting the English alphabetic characters in a string. In experiments we found that the Oblivious was easy for users, because human beings can recognize images of alphabetic characters quickly and accurately, but this is difficult for computers, because OCR tech-niques tend to misrecognize non-alphabetic characters as though they were alphabetic. This shows our approach is practical. We also describe novel algorithms for enhancing the skill gap between humans and computers that can be used with many existing CAPTCHAs.

抄録全体を表示

is a technology designed to prevent automated programs (known as bots) from acquiring access to on-line accounts to send spam mail, manipulate vote numbers in on-line polls, or take other malicious actions. In addition, access to Web services has been incorporated into mobile devices, such as smartphones. However, because most CAPTCHAs are not designed for mobile devices, user-friendly for mobile devices is required. Thus, we implemented on mobile devices and evaluated its resistance to bots. Our showed robustness against bots, with good usability.

抄録全体を表示

is designed to detect automated programs (called bots) by requiring them to perform tasks that are easy for humans but difficult for automations. CAPTCHAs are vulnerable to relay attacks in which the challenges are relayed to remote human-solvers. In our previous paper, we proposed an interactive video type that is strongly resistant to relay attacks. However, a quantitative evaluation of resistance to automated attacks still has not been carried out. Herein, we implement an automated attack for applying to our and evaluate its resistance to automated attacks. Our results show the robustness of our proposed method against mean shift algorithm.

抄録全体を表示

Although most existing text-based CAPTCHAs use distorted images of alphanumerics, they have been criticized because large image distortions make it difficult for human beings to recognize the characters, despite the ease with which computers can eliminate distortions and consequently recognize them. Ergonomically designed CAPTCHAs, which exploit human-specific phenomena, are a solution to this problem. Ergonomic design enables humans to momentarily recognize them, while significantly increasing computational costs for machines to recognize characters. Recently, owing to the development of deep learning, the image recognition capability of machines has improved dramatically. Characters are easily recognized within reasonable time by deep convolutional neural networks (DCNNs), which have similar architectures to visual perception mechanisms of the brain. In this paper, to clarify whether ergonomically designed CAPTCHAs can withstand state-of-the-art methods of deep learning, we use several kinds of DCNNs to measure the classification rates of the characters displayed. With respect to ergonomic designs, we first use Amodal

proposed by Mori et al., which exploits the two human-specific phenomena of amodal completion and aftereffects. We secondly modify Amodal by adding jagged lines to the edges of characters, aiming to prevent DCNNs from recognizing them correctly, since edges are one of the most fundamental features for DCNNs. Experimental results, however, show that both naive and jagged-lined Amodal CAPTCHAs are almost completely broken. Another approach we conducted is to use only complete characters without shielding as training data, assuming that attackers have no information about how amodal completion and jagged edges were applied. However, even for this assumption, the classification rate of DCNNs is still sufficiently high. On the whole, our results in this paper show that any ergonomic effects such as amodal completion and jagged edges are no longer countermeasures against character recognition by DCNNs.

抄録全体を表示

is a security system that is intended to distinguish computers and humans. Although many types of programs exist, many solver services are now becoming available. Humans can recognize partially-hidden objects that computers cannot. In this paper, we use this feature to propose a new text-based , in which sprinkled destructors are placed on and under the characters. Based on the evaluation of our method, we have found that a general character recognition software called Google Tesseract and a solver system are unable to recognize the generated by our method although humans can recognize them.

抄録全体を表示

近年, 既存のにおける脆弱性が多くの研究者によって指摘されており, 人間の「より高度な知識処理」に基づいた新たなの導入が強く望まれる.これに対し, 著者らは今までに, 「違和感を判別する能力」および「ユーモアを解する能力」という2つの人間の高度な認知処理に注目し, それぞれをチューリングテストに用いることで, 人間には容易で機械には困難な新しいを提案している.本稿ではそれら2つのについて紹介し, 利便性, 安全性, および運用性の観点から考察を加える.

抄録全体を表示

人間の視覚特性の一つである錯視を利用することでコンピュータによる解読を困難にしながらも, 人間には容易に解読できる文字型を作成する手法を提案する. 不規則的に配置した文字の背面と前面に任意形状の障害図形を配置することで, コンピュータによる解読を防ぐ. 提案手法の妥当性を検証するために行ったコンピュータと人間に対する認識実験の結果, コンピュータの認識率は0%, 人間の認識率は97%であった.

抄録全体を表示

人間の視覚特性の一つである錯視を利用することでコンピュータによる解読を困難にしながらも,人間には容易に解読できる文字型を作成する手法を提案する.不規則的に配置した文字の背面と前面に任意形状の障害図形を配置することで,コンピュータによる解読を防ぐ.提案手法の妥当性を検証するために行ったコンピュータと人間に対する認識実験の結果,コンピュータの認識率は0%,人間の認識率は97%であった.

抄録全体を表示

For the visually impaired, it is difficult to use

that was generated using visual information. The only way is to install a browser add-on and, by eliminating , to collect information from the internet. However, it is not easy for visually impaired people to install an add-on, and there will also be disadvantages for general users if anyone is able to eliminate the function. As such, using text that can be converted into voice, we proposed a sentence rearranging format of that makes use of the human ability to understand context. Through experiments, we examined the kind of sentences, number of Japanese characters, number of sentences, and special characteristics of sentences that the sentence rearranging format could accommodate. Our conclusion was that, primarily, humans find it easy to understand context to the extent that the subject of the sentence is clearly indicated and is linked to the content of the sentence, that it is appropriate to use four sentences or fewer as multiple-choice options in problem sentences, and that less than 100 Japanese characters is ideal for the number of characters.

抄録全体を表示

ウェブサイトへの迷惑メールの投稿の様な悪意を持った目的で， 人工的に合成された文章が多用されている．我々は，マルコフチェイン によって生成された文章と自然文章を区別する問題を研究する． この問題の困難性に基づいて，チャレンジレスポンス型の機械を 判別のテストであるへの応用を検討する．

抄録全体を表示

CAPTCHAs distinguish humans from automated programs by presenting questions that are easy for humans but difficult for computers, e.g., recognition of visual characters or audio utterances. The state of the art research suggests that the security of visual and audio CAPTCHAs mainly lies in anti-segmentation techniques, because individual symbol recognition after segmentation can be solved with a high success rate with certain machine learning algorithms. Thus, most recent commercial CAPTCHAs present continuous symbols to prevent automated segmentation. We propose a novel framework that can automatically decode continuous CAPTCHAs and assess its effectiveness with actual questions from Google's reCAPTCHA. Our framework is constructed on the basis of a sequence recognition method based on hidden Markov models (HMMs), which can be concisely implemented by using an off-the-shelf library HMM toolkit. This method concatenates several HMMs, each of which recognizes a symbol, to build a larger HMM that recognizes a question. Our experimental results reveal vulnerabilities in continuous CAPTCHAs because the solver cracks the visual and audio reCAPTCHA systems with 31.75% and 58.75% accuracy, respectively. We further propose guidelines to prevent possible attacking from HMM-based solvers on the basis of synthetic experiments with simulated continuous CAPTCHAs.

抄録全体を表示

久保田らは，タイポグリセミアと絵を用いたを提案しているが，ユーザに提示する設問は人手により作成されていた．本稿では，設問の自動作成方法を提案する．被験者50人を対象とした実験により，自動作成された設問は，人手により作成された設問と遜色ないことが確認された．

抄録全体を表示