|
所属 |
工学教育研究部 工学科情報通信プログラム担当 |
|
職名 |
特別助教 |
|
関連SDGs |
|
論文 【 表示 ／ 非表示 】
-
画像内の反射成分の干渉下でも検出性能を維持できる画像中のプライバシーリスク検出手法 査読あり
臼崎,翔太郎, 油田,健太郎, 山場,久昭, 朴,美娘, 岡崎,直宣
情報処理学会論文誌 67 ( 3 ) 588 - 604 2026年3月
記述言語：日本語 掲載種別：研究論文（学術雑誌） 出版者・発行元：情報処理学会
本研究では，SNSでの画像投稿時の利用を想定し，画像内の反射成分による干渉下でも検出性能を維持できる，画像中のプライバシーリスク検出手法を提案する．反射成分とは，画像内において，透過物質越しに撮影した際に透過物質の手前の風景が反射し，本来の撮影対象に重なって映っている成分を指す．反射成分の像が重なって映るため，反射成分を含まない画像に比べてリスクの検出がしにくいことが予想される．本研究では，画像中に反射成分と透過成分が存在することを前提とし，次の3つの工夫を施したモデルを提案する．(1)反射成分の物体検出用の出力を追加する，(2)検出タスクだけでなく反射成分の分離タスクの同時学習を行う，(3)分離タスクの推測結果を検出タスクの補助情報として利用する．実験では，反射成分が存在する疑似的な画像を利用し，提案モデルと類似手法，個別に学習した手法と検出精度を比較した．実験の結果，提案モデルは反射成分が存在する画像でのリスク検出精度を一般的な手法よりも改善できた．検出精度をより高めるには，適切なパラメータ調整や学習方法を検討する必要があることが分かった．
In this study, we propose a novel image privacy risk detection method that accounts for an image's reflection component. The reflection component refers to the undesired entered component in which objects ahead of a transparent surface (e.g., glass) are reflected and overlapped onto the subject beyond the transparent surface. Since the transmission and reflection components spatially overlap in the image, detecting privacy risks becomes more challenging compared to images without such reflection. Assuming that both reflected and transmitted components coexist in an image, we propose a privacy risk detection model that incorporates the following three design elements: To this end, our approach introduced the following three strategies: (1) extending the output of the detection decoder with additional channels dedicated to reflection components; (2) formulating the problem as a multitask learning that simultaneously optimizes the privacy risk detection task and the reflection removal task; and (3) utilizing the estimated reflection components as auxiliary features for the detection task by channel attention. We conducted experiments using synthetically generated images that include reflection components and evaluated the performance of our proposed method against baseline models, including those trained on each task individually. The results demonstrate that the proposed method achieves superior detection accuracy compared to general approaches, confirming its effectiveness in handling privacy risks by reflection. It became clear that adjusting the appropriate parameters and learning methods were necessary to improve the detection performance.DOI： 10.20729/0002008648
-
2010年宮崎県口蹄疫を事例とした感染リスク評価指標の構築と防疫支援への応用 査読あり
高塚 佳代子, 臼崎 翔太郎, 山場 久昭, 油田 健太郎, 岡崎 直宣
化学工学論文集 52 ( 1 ) 15 - 20 2026年1月
記述言語：日本語 掲載種別：研究論文（学術雑誌） 出版者・発行元：公益社団法人 化学工学会
畜産業における伝染病の蔓延は，深刻な経済的損失と社会的混乱を引き起こす．本研究は，2010年に宮崎県で発生した口蹄疫の事例を基に，農場間の空間的関係，飼養規模，動物種といった複数要因を統合的に考慮した「潜在的感染リスク指数（Potential Infection Risk Index: λ）」を構築し，感染の有無にかかわらず農場単位でのリスクを定量的に評価する手法を提案する．統計分析により，λは感染農場と未感染農場の間で有意差を示し，感染源からの距離に応じてその差異が拡大する傾向が確認された．これにより，感染の広がりは地理的近接性に加えて農場の構造的特性に強く依存することが示唆された．さらに，λを用いて感染リスクの高い農場を客観的かつ定量的に特定でき，防疫資源の優先的配分や飼養衛生管理基準の柔軟な運用といった実践的課題への対応が可能となる．加えて，農場移転を想定したケーススタディでは，λの低下によりリスク軽減効果が確認され，λが防疫的・経済的に合理的な意思決定支援指標として有用であることが実証された．今後は，λの可視化・操作性を備えた支援ツールとしての実装を進めることで，防疫判断の迅速化・高度化が期待される．
-
Usuzaki S., Iwasa K., Takahashi N., Aburada K., Yamaba H., Park M., Okazaki N.
Proceedings 9th International Conference on Information Technology Incit 2025 545 - 552 2025年11月
掲載種別：研究論文（国際会議プロシーディングス） 出版者・発行元：Proceedings 9th International Conference on Information Technology Incit 2025
In this paper, we discuss a sketch-reconstructionbased In-band Network Telemetry (INT) method that can achieve practical flow count estimation accuracy during sketch collection. Measuring and aggregating per-flow information of network traffic helps to monitor the current network-wide status. Sketch-reconstruction-based INT methods were proposed to efficiently collect flow information measured by each device, which combine sketch techniques and INT. Sketch is a probabilistic algorithm that can estimate the count of entire flows with limited memory, and INT can carry measured data on the network device piggybacked in the business packet header. The combination of sketch and INT can reduce bandwidth consumption for collecting per-flow information, as the sketch can be regarded as a compact aggregation of the per-flow data. To minimize overhead, these methods split a sketch into smaller units called sketchlets and transmit them. The sent sketchlets are reconstructed into a sketch at the collection destination. However, existing methods have the disadvantage that they require a long time to collect sketchlets because they randomly select them, and the accuracy is low during the collection. In this study, we propose a new sketchlet algorithm, called HitFlow Sketchlet, that prioritizes sending buckets necessary for the sketch's query algorithm to function, which each method does not fully consider. We also propose the Flag data structure, aiming at the comprehensive sketchlet transmission. Evaluation experiments showed that HitFlow Sketchlet can output accurate estimates, especially in the earliest epochs, and that by combining it with Flag, comprehensiveness can be improved.
-
Conjugate Gradient Based Multiplicative Update Rules for Nonnegative Matrix Factorization 査読あり
Takahashi N., Usuzaki S., Katayama T., Yokomichi M., Aburada K., Okazaki N.
Proceedings 9th International Conference on Information Technology Incit 2025 639 - 644 2025年11月
掲載種別：研究論文（国際会議プロシーディングス） 出版者・発行元：Proceedings 9th International Conference on Information Technology Incit 2025
In this paper, we propose a general solution for nonnegative matrix factorization (NMF), a technique widely used in the field of machine learning. NMF is a type of matrix factorization used for dimensionality reduction. By approximating a nonnegative matrix as the product of two nonnegative matrices, it is possible to extract frequent patterns in the data matrix into the basis matrix together with their corresponding weights. Since many types of information, such as images, spectrograms of acoustic signals, and text data, are represented by nonnegative values, NMF is a useful algorithm applicable across a wide range of fields. Standard approaches to solving NMF include algorithms based on multiplicative update rules. However, because the applicable algorithm depends on the problem structure, extending NMF to new problem settings can be challenging. Another approach is the gradient method, which is an additive iterative procedure for solving a nonlinear optimization problem. The update direction of the state is given by the gradient of the objective function. The gradient method is versatile. It can be applied to any optimization problem with an available gradient, regardless of the problem structure. However, when applying it to the NMF problem, it is necessary to maintain the nonnegativity of the matrices throughout the iterative computation. As a result, several issues must be addressed, including vanishing gradients, proper tuning of the learning rate, and the need for gradient clipping. In this paper, we propose a new algorithm based on a multiplicative update rule incorporating the gradient. A multiplier function consisting of the hyperbolic tangent function and the linear functions is used. This function ensures that the pair of factor matrices remain nonnegative throughout the entire computation process. As with general gradient methods, the proposed algorithm has the advantage of being widely applicable to NMF problems formulated in a differentiable form. Furthermore, the convergence speed is improved by utilizing the nonlinear conjugate gradient method. Numerical experiments show that the decomposed matrices have a property specified by the penalty terms and confirming the effectiveness of the proposed method.
-
Yamaba H., Komori N., Usuzaki S., Aburada K., Katayama T., Okazaki N.
Icce Taiwan 2025 12th IEEE International Conference on Consumer Electronics Taiwan Generative AI in Innovative Consumer Technology Proceedings 111 - 112 2025年
掲載種別：研究論文（学術雑誌） 出版者・発行元：Icce Taiwan 2025 12th IEEE International Conference on Consumer Electronics Taiwan Generative AI in Innovative Consumer Technology Proceedings
Mobile devices such as smartphones and tablets are now deeply integrated into daily life. To prevent shoulder surfing attacks, we propose a user authentication method for mobile devices based on surface electromyogram (s-EMG) signals. The method uses a “pass-gesture,” a sequence of hand gestures that can be changed like a password. Fingerspelling was adopted as the source of gesture candidates. Previous studies confirmed high accuracy in identifying gestures performed by the same subject. This paper focuses on evaluating subject specificity-specifically, confirming that a model trained on one user fails to recognize gestures performed by another. Experimental results show that the system does not generalize across users, indicating robustness against impersonation.
書籍等出版物 【 表示 ／ 非表示 】
-
Slow HTTP POST DDoS Attack Prevention Method that Monitors Payload Size and Number of Connections
Ozaki Y., Usuzaki S., Aburada K., Yamaba H., Park M., Okazaki N.（ 担当： 単著）
Lecture Notes on Data Engineering and Communications Technologies 2026年
著書種別：学術書
The threat of distributed denial-of-service (DDoS) attacks has increased in recent years. In particular, the slow HTTP POST attack overwhelms web servers by maintaining slow connections with small payloads. Previous research has addressed this threat by monitoring the number of simultaneous connections and using programmable P4 switches. However, these approaches have shown limitations when facing distributed attacks from multiple IP addresses. In this study, we developed an improved defense method that monitors both the number of simultaneous connections and the payload size to more effectively identify a slow HTTP POST DDoS attack by capturing its characteristics. Experimental comparisons with existing methods demonstrated that connection-based monitoring alone is insufficient against distributed attacks, whereas our new method achieves better performance. Ultimately, implementing this method on a P4 switch is expected to reduce server load by blocking malicious traffic before it reaches the server, thereby enhancing the robustness of the entire system.
MISC 【 表示 ／ 非表示 】
-
An Attempt to Solve Fill in the Missing Letters CAPTCHA Using Generative AI 査読あり
Yamaba H., Usuzaki S., Aburada K., Mukunoki M., Park M., Okazaki N.
Lecture Notes in Electrical Engineering 1322 LNEE 385 - 395 2024年
掲載種別：速報，短報，研究ノート等（学術雑誌） 出版者・発行元：Lecture Notes in Electrical Engineering
This paper reports an attempt to solve the fill in the missing letters type CAPTCHA using generative AI. Many websites have adopted CAPTCHA to prevent bots and other automated programs from engaging in malicious activities such as posting comment spam. Text-based CAPTCHA is the most common and earliest form of CAPTCHA. However, as optical character recognition (OCR) technology has improved, the intensity of distortions applied to a CAPTCHA to keep it unrecognizable by OCR has also increased. This has reached a point where humans are having difficulty recognizing CAPTCHA text. The CAPTCHA proposed in the previous study asks users to spell a word by filling in some blanks. Since the number of letters displayed is minimal, it is challenging to identify the correct word. However, one or more images that can serve as hints to help users guess the answer word are also provided. It is expected that the ability to guess can distinguish between humans and computers. However, it is conceivable that generative AI, which has been advancing in recent years, can substitute for this ability. A series of experiments was carried out to evaluated the performance of the generative AI’s ability to solve the proposed CAPTCHA. First, we examined whether a well-known image recognition system could accurately identify the images used in the CAPTCHA problems. Next, we used the recognition results to have the generative AI solve the CAPTCHA problems and determined the accuracy rate. Additionally, we evaluated the performance of the generative AI itself by solving the problems using the correct identification of each image. From the experimental results, it was found that the CAPTCHA is relatively robust against attack techniques using generative AI.
DOI： 10.1007/978-981-96-1535-3_38
その他リンク： https://dblp.uni-trier.de/db/conf/icgec/icgec2024-2.html#YamabaUAMPO24
-
Park M., Chang Z., Usuzaki S., Aburada K., Okazaki N.
Proceedings 2024 12th International Symposium on Computing and Networking Workshops Candarw 2024 293 - 299 2024年
掲載種別：速報，短報，研究ノート等（学術雑誌） 出版者・発行元：Proceedings 2024 12th International Symposium on Computing and Networking Workshops Candarw 2024
In recent years, with the number of users utilizing online services increasing, there is growing demand for more secure authentication methods. Existing authentication methods such as passwords, PINs, and pattern locks, while highly operational and convenient, have issues such as vulnerability to brute force attacks and voyeuristic attacks. As a solution to these problems, gaze authentication has been proposed, where a user draws pre-defined characters or symbols with their gaze, and the features obtained from the trajectory information are used for personal identification. Gaze-based authentication is resistant to attacks such as voyeuristic attacks, thermal attacks, and smudge attacks because it is difficult for a third party to observe the authentication process. However, it also has drawbacks, such as a low authentication success rate and a long authentication time. Therefore, this study proposes a personal authentication system that combines gaze-based image selection and gaze trajectory features to address the issues of authentication rate and authentication time in gaze-based authentication. Additionally, to verify the effectiveness of the proposed system, we collected eye movement trajectory data from users using a gaze detection device, performed data preprocessing and feature extraction, and then calculated the personal authentication rate using several machine learning algorithms. We confirmed that XGBoost had the highest authentication accuracy with an F-measure of 67.5%. Furthermore, in the evaluation of anomaly detection algorithms, Isolation Forest demonstrated superior performance compared to other algorithms.
DOI： 10.1109/CANDARW64572.2024.00055
その他リンク： https://dblp.uni-trier.de/db/conf/candar/candar2024w.html#ParkCUAO24
-
Evacuation Support System for Tsunami Disasters that Considers Shelter Congestion 査読あり
Nishi H., Nabeyama K., Usuzaki S., Aburada K., Yamaba H., Katayama T., Park M., Okazaki N.
Lecture Notes in Electrical Engineering 1322 LNEE 51 - 62 2024年
掲載種別：速報，短報，研究ノート等（学術雑誌） 出版者・発行元：Lecture Notes in Electrical Engineering
When a tsunami occurs, an evacuation support system capable of providing information on evacuation routes and shelter locations would enable evacuations to be conducted more swiftly and safely. In previous research, a proposal was made for an evacuation support system using low-power, long-distance communication. However, a challenge arose when shelter locations reached full capacity, requiring information to be shared, which in turn necessitated redirection of evacuees to alternative shelters. Therefore, in this study, a method for sharing shelter congestion information at an earlier stage was developed. The method of sharing based on the remaining capacity of shelters increased evacuation completion by up to 29 evacuees in an evacuee population of 1000.
DOI： 10.1007/978-981-96-1535-3_7
その他リンク： https://dblp.uni-trier.de/db/conf/icgec/icgec2024-2.html#NishiNUAYKPO24
-
Usuzaki S., Aburada K., Yamaba H., Katayama T., Mukunoki M., Park M., Okazaki N.
Artificial Life and Robotics 27 ( 1 ) 2022年2月
掲載種別：速報，短報，研究ノート等（学術雑誌） 出版者・発行元：Artificial Life and Robotics
In the original publication of the article, on Figs. 4 and 5, “CNN” should read as “MLP”. In addition, under the section “4.2 Human and machine success rate”, on the paragraph “For evaluation of the machine success …”, in the following sentence “In the attack experiment, we applied the famous color …” CNN (Convolutional Neural Network) should be corrected as MLP (Multilayer perceptron). The correct sentence should read as “In the attack experiment, we applied the famous color constancy algorithm, Gray-World, Max-RGB, Gray-Edge [11], Second Derivative Gray-Edge, and MLP (Multilayer perceptron) to images that were saved in the experiment for the human success rate
科研費（文科省・学振・厚労省）獲得実績 【 表示 ／ 非表示 】
-
深層学習を用いた筋電位による多要素個人認証の複数点測定による高精度化
研究課題/領域番号：24K14948 2024年04月 - 2027年03月
独立行政法人日本学術振興会 科学研究費基金 基盤研究(C)
担当区分：研究分担者
-
AIによって高度化されたサイバー攻撃に対抗する次世代クラウドセキュリティ基盤の構築
研究課題/領域番号：24K14917 2024年04月 - 2027年03月
独立行政法人日本学術振興会 科学研究費基金 基盤研究(C)
担当区分：研究分担者
-
透過的反射による人間の映り込みを高精度に検出するSNS投稿画像解析手法の開発
研究課題/領域番号：23H05387 2023年04月 - 2024年03月
独立行政法人日本学術振興会 科学研究費補助金 奨励研究
担当区分：研究代表者