Yuko Hara-Azumi

Yuko Hara-Azumi

Auto-generated profile, based on publicly available information

Yuko Hara-Azumi

Institute of Science Tokyo

Research Focus & Metrics

Academic Career Length

15Years (2012-Present)

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

Representative Publications

View All Publications
7 September 2020 in IEEE Access
Verified
Journal Article
Open access
PAPaniti Achararit
Paniti Achararit
MHMuhammad Abdullah Hanif
Muhammad Abdullah Hanif
RPRachmad Vidya Wicaksana Putra
Rachmad Vidya Wicaksana Putra
DSDR SHAFIQUE
DR SHAFIQUE
YHYuko Hara-Azumi
Yuko Hara-Azumi
Designing resource-efficient deep neural networks (DNNs) is a challenging task due to the enormous diversity of applications as well as their time-consuming design, training, optimization, and evaluation cycles, especially the resource-constrained embedded systems. To address these challenges, we propose a novel DNN design framework called accuracy-and-performance-aware neural architecture search (APNAS), which can generate DNNs efficiently, as it does not require hardware devices or simulators while searching for optimized DNN model configurations that offer both inference accuracy and high execution performance. In addition, to accelerate the process of DNN generation, APNAS is built on a weight sharing and reinforcement learning-based exploration methodology, which is composed of a recurrent neural network controller as its core to generate sample DNN configurations. The reward in reinforcement learning is formulated as a configurable function to consider the sample DNNs’ accuracy and cycle count required to run on a target hardware architecture. To further expedite the DNN generation process, we devise analytical models for cycle count estimation instead of running millions of DNN configurations on real hardware. We demonstrate that these analytical models are highly accurate and provide cycle count estimates identical to those of a cycle-accurate hardware simulator. Experiments that involve quantitatively varying hardware constraints demonstrate that APNAS requires only 0.55 graphics processing unit (GPU) days on a single Nvidia GTX 1080Ti GPU to generate DNNs that offer an average of 53% fewer cycles with negligible accuracy degradation (on average 3%) for image classification compared to state-of-the-art techniques.

18 January 2024 in IEEE Internet of Things Journal
Verified
Journal Article
Open access
MYMingyu Yang
Mingyu Yang
TATanvir Ahmed
Tanvir Ahmed
SISaya Inagaki
Saya Inagaki
KSKazuo Sakiyama
Kazuo Sakiyama
YLYang Li
Yang Li
YHYuko Hara-Azumi
Yuko Hara-Azumi
With the growth of Internet of Things (IoT) era, the protection of secret information on IoT devices is becoming increasingly important. For IoT devices, attacks that target information leakage through physical side-channels (e.g., a power side-channel) are a major threat in many use cases because IoT devices can be accessed easily by a hostile third party. However, securing resource-constrained IoT devices against side-channel attacks is a challenging issue. Generally, it is difficult to satisfy the requirements on side-channel protection while maintaining the low-power and real-time constrains of IoT devices. In this paper, we propose a hardware/software cooperative design for cryptosystems that is suitable for resource-constrained IoT devices. Combining a security-oriented processor design (i.e., an instruction set architecture definition and its architectural structure) and careful implementations of masked software implementation for cipher algorithms can effectively improve the power-performance-area (PPA) while suppressing power side-channel leakage. In our evaluation, for three ciphers (Chaskey, Simon, and AES), we demonstrate that our work is superior to state-of-the-art works (two RISC-V processors and a small-scale low-power processor) in terms of both PPA and power side-channel protection.

1 January 2016
Verified
Conference Paper
Closed access
PDF file
JAJason H. Anderson
Jason H. Anderson
YHYuko Hara-Azumi
Yuko Hara-Azumi
SYShigeru Yamashita
Shigeru Yamashita
Stochastic computing (SC) [1] has received attention recently as a paradigm to improve energy efficiency and fault tolerance. SC uses hardware-generated random bitstreams to represent numbers in the [0:1] range - the number represented is the probability of a bit in the stream being logic-1. The generation of random bitstreams is typically done using linear-feedback shift register (LFSR)-based random number generators. In this paper, we consider how best to design such LFSR-based stochastic bitstream generators, as a means of improving the accuracy of stochastic computing. Three design criteria are evaluated: 1) LFSR seed selection, 2) the utility of scrambling LFSR output bits, and 3) the LFSR polynomials (i.e. locations of the feedback taps) and whether they should be unique vs. uniform across stream generators. For a recently proposed multiplexer-based stochastic logic architecture [8], we demonstrate that careful seed selection can improve accuracy results vs. the use of arbitrarily selected seeds. For example, we show that stochastic logic with seed-optimized 255-bit stream lengths achieves accuracy better than that of using 1023-bit stream lengths with arbitrary seeds: an improvement of over 4 in energy for equivalent accuracy.

20 January 2022 in IEEE Access
Verified
Journal Article
Open access
MAMehdi Trabelsi Ajili
Mehdi Trabelsi Ajili
YHYuko Hara-Azumi
Yuko Hara-Azumi
Internet of Things and deep learning (DL) are merging into one domain and enabling outstanding technologies for various classification tasks. Such technologies are based on complex networks that mainly target powerful platforms with rich computing resources, such as servers. Therefore, for resource-constrained embedded systems, new challenges of size, performance (i.e., latency, throughput, and accuracy), and power consumption-aware networks should be addressed, particularly when edge devices handle multimodal data (i.e., different types of real-time sensing data). In this case study, we focus on DeepSense, a time-series multimodal DL framework combining convolutional neural networks (NN) and recurrent NN to process accelerometer and gyroscope data for human activity recognition. We present a field-programmable gate array (FPGA)-based acceleration for DeepSense incorporated into a hardware/software co-design approach to achieve better latency and energy efficiency using the Xilinx Vitis AI framework. The architecture of DeepSense has drawbacks that cannot be easily alleviated by Vitis AI; therefore, we introduced a new methodology of adjusting the framework and its components (i.e., the deep learning processing unit (DPU)) to achieve a custom design suitable for such time-series multimodal NN. We implemented the accelerator on two FPGA boards and performed a quantitative evaluation by varying the DPU parameter settings to support our design approach. We demonstrated the effectiveness of our implementation against the original software implementation on mobile devices by achieving up to $2.5\times $ and $5.2\times $ improvement in latency and energy consumption, respectively. Through this case study, we provide crucial insights into the FPGA-based accelerator design of multimodal NN and essential aspects to consider for further improvements and adaptation in other application domains.

16 August 2021 in IEEE Access
Verified
Journal Article
Open access
YNYu Nakayama
Yu Nakayama
RYRyo Yaegashi
Ryo Yaegashi
ANAnh Hoang Ngoc Nguyen
Anh Hoang Ngoc Nguyen
YHYuko Hara-Azumi
Yuko Hara-Azumi
Many applications such as Industrial Internet of Things (IIoT), tactile Internet, and 5G/6G mobile require Ultra-Low Latency (ULL) in data transmission. The end-to-end latency is required to be on the order of sub-millisecond in these applications. The IEEE 802.1 Time Sensitive Networking (TSN) standards have been developed to provide ULL networking. The IEEE 802.1Qbv Time-Aware Shaper (TAS), which is a typical flow control mechanism of TSN, is a traffic shaper to provide deterministic end-to-end ULL transmission for express traffic. The Gate Control List (GCL) of all TAS enabled nodes must be configured in a coordinated manner to ensure ULL. However, existing configuration schemes cannot be employed in highly dynamic conditions where the distribution of time-sensitive streams frequently changes. It will be a significant problem to deal with such dynamic conditions in the future. Therefore, this paper proposes a real-time adaptive gate scheduling scheme for TAS. The scheduling problem is formulated as a boolean satisfiability problem (SAT), and the employment of an FPGA-based solver is proposed to achieve runtime fast computation. The proposed scheme enables real-time reconfiguration for high flexibility and high bandwidth utilization to deal with the dynamics of network conditions. The feasibility of the proposed scheme is confirmed with computer simulations.

22 March 2024 in IEEE Internet of Things Journal
Verified
Journal Article
Open access
QZQingyu Zeng
Qingyu Zeng
YHYuko Hara-Azumi
Yuko Hara-Azumi
The rapid expansion of the Internet of Things (IoT) has increased security concerns, thereby necessitating efficient intrusion detection systems (IDSs). In this article, we propose a real-time IoT IDS designed by combining a random forest (RF) classifier with an ensemble feature selection technique (EFST). The proposed IDS can be deployed on a small-scale field-programmable gate array (FPGA) board. The system utilizes a two-metric ensemble feature selection process to reduce computational complexity and enhance classification accuracy. In addition, the EFST aggressively extracts a limited number of features, thereby reducing the complexity of the RF model. Then, the tailored RF classifier is mapped onto an FPGA-based hardware accelerator to realize real-time detection. The proposed method was evaluated experimentally on the benchmark BoT-IoT data set. The results demonstrate that the proposed IDS realizes significant improvements in terms of resource utilization and processing time compared to several state-of-the-art FPGA-based IDS implementations while maintaining sufficient detection accuracy. In particular, our implementation on the Xilinx PYNQ Z2 achieved $10.2\times $ , $135.7\times $ , and $8.43\times $ speed-up compared to state-of-the-art IDSs running on an Intel Core i7 CPU, an ARM Cortex-A9 microprocessor, and a neural network-based accelerator on the PYNQ, respectively. In addition, our approach exhibits the lowest resource utilization among FPGA-based IDS solutions. These results demonstrate that this work contributes to developing secure and sustainable IoT ecosystems by integrating EFST, RF classification, and FPGA-based acceleration.