Kota Yoshida

Dr. Kota Yoshida

orcid

Research Focus & Metrics

Academic Career Length

6Years (2021-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
1 November 2021 in Microprocessors and Microsystems
Verified
Journal Article
Closed access
TKTakaya Kubota
Takaya Kubota
KYKota Yoshida
Kota Yoshida
MSMitsuru Shiozaki
Mitsuru Shiozaki
TFTakeshi Fujino
Takeshi Fujino

3 October 2024 in Sensors
Verified
Journal Article
Open access
PDF file
TITakahito Ino
Takahito Ino
KYKota Yoshida
Kota Yoshida
HMHiroki Matsutani
Hiroki Matsutani
TFTakeshi Fujino
Takeshi Fujino
In this paper, we introduce a security approach for on-device learning Edge AIs designed to detect abnormal conditions in factory machines. Since Edge AIs are easily accessible by an attacker physically, there are security risks due to physical attacks. In particular, there is a concern that the attacker may tamper with the training data of the on-device learning Edge AIs to degrade the task accuracy. Few risk assessments have been reported. It is important to understand these security risks before considering countermeasures. In this paper, we demonstrate a data poisoning attack against an on-device learning Edge AI. Our attack target is an on-device learning anomaly detection system. The system adopts MEMS accelerometers to measure the vibration of factory machines and detect anomalies. The anomaly detector also adopts a concept drift detection algorithm and multiple models to accommodate multiple normal patterns. For the attack, we used a method in which measurements are tampered with by exposing the MEMS accelerometer to acoustic waves of a specific frequency. The acceleration data falsified by this method were trained on an anomaly detector, and the result was that the abnormal state could not be detected.

27 March 2023 in Journal of Cryptographic Engineering
Verified
Journal Article
Open access
PDF file
KKKunihiro Kuroda
Kunihiro Kuroda
YFYuta Fukuda
Yuta Fukuda
KYKota Yoshida
Kota Yoshida
TFTakeshi Fujino
Takeshi Fujino
Deep-learning side-channel attacks, applying deep neural networks to side-channel attacks, are known that can easily attack some existing side-channel attack countermeasures such as masking and random jitter. While there have been many studies on profiled deep-learning side-channel attacks, a new approach that involves applying deep learning to non-profiled attacks was proposed in 2018. In our study, we investigate the structure of multi-layer perceptrons and points of interest for non-profiled deep-learning side-channel attacks using the ANSSI database with a masking countermeasure. The results of investigations indicate that it is better to use a simple network model, apply regularization to prevent over-fitting, and select a wide range of power traces that contain side-channel information as the points of interest. We also implemented AES-128 software implementation protected with the Rotating Sboxes Masking countermeasure, which has never been attacked by non-profiled deep-learning side-channel attacks, on the Xmega128 microcontroller and carried out non-profiled deep-learning side-channel attacks against it. Non-profiled deep-learning side-channel attacks successfully recovered all partial keys while the conventional power analysis could not. The attack results also showed that the least significant bit is the adequate selection for successful non-profiled deep-learning side-channel attacks, but the best labeling method may vary depending on the implementation of the countermeasure algorithm. We conducted two experimental analyses to clarify that deep-learning side-channel attacks learn mask values used in the masking countermeasure. One is the gradient visualization used in previous studies, and the other is a new analysis method using partial removal of power traces.

31 October 2024 in Electronics
Verified
Journal Article
Open access
PDF file
YMYudai Morikaku
Yudai Morikaku
RURyuichi Ujiie
Ryuichi Ujiie
DMDaisuke Morikawa
Daisuke Morikawa
HSHideki Shima
Hideki Shima
KYKota Yoshida
Kota Yoshida
SOShunsuke Okura
Shunsuke Okura
In order to improve image recognition technologies in an IoT environment, we propose a data reduction scheme for a feature-extractable CMOS image sensor and present simulation results for object recognition using feature data. We evaluated the accuracy of the simulated feature data in object recognition based on YOLOX trained with a feature dataset. According to our simulation results, the obtained object recognition accuracy was 56.6% with the large-scale COCO dataset, even though the amount of data was reduced by 97.7% compared to conventional RGB color images. When the dataset was replaced with the RAISE RAW image dataset for more accurate simulation, the object recognition accuracy improved to 76.3%. Furthermore, the feature-extractable CMOS image sensor can switch its operation mode between RGB color image mode and feature data mode. When the trigger for switching from feature data mode to RGB color image mode was set to the detection of a large-sized person, the feature data achieved an accuracy of 93.5% with the COCO dataset.

27 November 2024 in IEEE Open Journal of Intelligent Transportation Systems
Verified
Journal Article
Open access
MOMasato Okuda
Masato Okuda
KYKota Yoshida
Kota Yoshida
TFTakeshi Fujino
Takeshi Fujino
In recent intelligent transportation systems (ITS), it is important to recognize pedestrians and avoid collisions. Various sensors are used to detect pedestrians, and some research on pedestrian detection uses a visible light (RGB) camera and a far-infrared (FIR) camera. FIR cameras are significantly affected by ambient temperatures such as summer and winter. However, few studies have focused on this property when evaluating pedestrian detection accuracy. Therefore, this paper investigates the effect of temperature change in real-time multispectral pedestrian detection. We created an original dataset with three subsets, Hot, Intermediate, and Cold, and evaluated temperature effects by changing the subsets during training and testing. We first evaluated YOLOv8s-4ch, which simply extended the input layer of YOLOv8 from 3 channels of RGB to 4 channels of RGB-FIR. To further improve detection performance, we built a new model called YOLOv8s-2stream. This model has two backbones for RGB and FIR, and fuses their feature maps in each resolution. We found that the model trained on a specific temperature subset dropped the test accuracy in other subsets. On the other hand, when training using a Mix set covering all temperature sets (Hot, Inter., Cold), the model achieved the highest accuracy through all conditions. Moreover, our YOLOv8s-2stream has improved by 3.9 points of accuracy (AP@0.5:0.95) compared to YOLOv8s-4ch, and achieved 73FPS inference speed on Jetson.

14 May 2023 in Sensors
Verified
Journal Article
Open access
PDF file
TOTatsuya Oyama
Tatsuya Oyama
SOShunsuke Okura
Shunsuke Okura
KYKota Yoshida
Kota Yoshida
TFTakeshi Fujino
Takeshi Fujino
A backdoor attack is a type of attack method that induces deep neural network (DNN) misclassification. The adversary who aims to trigger the backdoor attack inputs the image with a specific pattern (the adversarial mark) into the DNN model (backdoor model). In general, the adversary mark is created on the physical object input to an image by capturing a photo. With this conventional method, the success of the backdoor attack is not stable because the size and position change depending on the shooting environment. So far, we have proposed a method of creating an adversarial mark for triggering backdoor attacks by means of a fault injection attack on the mobile industry processor interface (MIPI), which is the image sensor interface. We propose the image tampering model, with which the adversarial mark can be generated in the actual fault injection to create the adversarial mark pattern. Then, the backdoor model was trained with poison data images, which the proposed simulation model created. We conducted a backdoor attack experiment using a backdoor model trained on a dataset containing 5% poison data. The clean data accuracy in normal operation was 91%; nevertheless, the attack success rate with fault injection was 83%.