Secure the Internet of Things (IoT) using Physical Layer Security and Blockchain

Hi, everyone! This is Yuanyu Zhang from the Large-Scale Systems Management Laboratory.  It is a great opportunity for me to share my research with you. Broadly speaking, my research focuses on the security of Internet of Things (IoT) systems, with special interests in securing IoT wireless communications based on physical layer security (PLS) technology and designing access control schemes based on the emerging blockchain technology. Now, I am going to briefly introduce my research from these two aspects.

  • PLS-based Secure IoT Wireless Communications 

When it comes to the IoT, I believe anyone can give an concrete application, such as smart home, e-health, intelligent transportation, etc. In these applications, a huge number of smart devices, like sensors, actuators, tablets, will be connected into the Internet via a variety of wireless communication technologies, like WiFi, Zigbee and Bluetooth.  As we all know, wireless medium is open such that anyone close to a transmitter can receive its signals. Suppose the information conveyed by the signal is not encrypted, so what would happen if the signal is received by an eavesdropper? Obviously, the information is in danger! In IoT applications, these information may be your health data perceived by sensors around you and your financial transfer records sent from smart phones. Once these information are leaked to eavesdroppers, your life and property will be in danger.

Of course, we can encrypt these information using secret key-based cryptographic approaches like what we do in wired communications. However, managing and distributing secret keys in wireless environment is challenging, especially for large-scale IoT systems. Besides, the cryptographic approaches usually require high computing power, which is unavailable for most resource-constrained IoT devices. This motivates the advent of the so-called physical layer security (PLS) technology, which uses the inherent randomness of wireless channels (e.g., fading, noise, interference) instead of costly cryptographic approaches to ensure no information leaked to eavesdroppers. As long as we can guarantee that the received signal of the eavesdropper is an degraded version of that of the intended receiver, the eavesdropper can abstract nothing from the signal. This conclusion has been proved from the perspective of information theory. The figures below shows the basic differences between cryptographic approaches and PLS technology.

What my research does is to apply commonly-used PLS techniques in IoT systems and theoretically evaluate the security performances of these systems using some mathematical tools, like Probability Theory, Markov Chain Theory, Queuing Theory and Stochastic Geometry.  Well, okay, I know it’s boring, but trust me, it is vitally important for understanding the PLS performance limits of IoT systems and contributing to the successful application of PLS techniques in these systems. If it does not sound boring to you, I would be very glad to share this research with you in greater details.

  • Blockchain-based IoT Access Control

Okay, ready for another research? You know what, this research may be more interesting, at least for me. The previous research focuses on the security of the information during its transmission, while this research focuses on the security of the information when it is stored somewhere as resource and accessed by some users or devices. Of course, resources are not limited to information, but also include actuators that may perform some critical tasks, like brake control and drug delivery.  But one common thing is that, once these resources are accessed by malicious users, your life and property will be in danger. This research aims to design effective access control schemes to prevent illegal access to IoT resources.

Traditional access control schemes are centralized, which means that they use a central server to control all access requests in the system. So what if this server is destroyed in disasters, or what if this server is compromised by some malicious guy? Yes, the whole access control scheme collapses.  Now, question is how to tackle these challenging issues. Fortunately, the emerging blockchain technology provides us with a promising solution, because it is highly distributed and ensures reliable financial transactions among trustless peers all over the world. In this research, I use the Ethereum blockchain, which evolves into a distributed and reliable computing platform thanks to the introduction of smart contract. A smart contract can be thought of as a piece of code that is stored on the blockchain and distributed to all nodes in the system. All nodes can execute this code and verify the correctness of the results. This ensures the correct execution of the smart contract as long as no one possesses more than 51% of the system computing power.

The basic idea of this research is to ensure  distributed and reliable access control by expressing the access control policies and logic as smart contracts. The access control framework is illustrated in the figure below.

This framework consists of multiple access control contracts (ACCs), one judge contract (JC) and one register contract (RC). Each ACC is responsible for the access control of a subject-object pair, and implements both static access right validation based on predefined policies and dynamic access right validation by checking the behavior of the subject. Here, a subject is the node accessing the resources possessed by an object. The JC implements the misbehavior-judging method to facilitate the dynamic validation of the ACCs by receiving misbehavior reports from the ACCs, judging the misbehavior and returning the corresponding penalty. The RC registers the information of ACCs and JC, and also provides functions (e.g., register, update and delete) to manage these methods. Suppose a server wants to access the resource of a camera in a smart home. The workflow of the access control is shown in the figure below.

  • Step 1: The server calls the RC to retrieve the ACC (e.g., the ACC 2) for access control.
  • Step 2: The RC returns the address and ABI (similar to API) of the ACC to the server.
  • Step 3: The server sends a transaction to the ACC, which contains the required information for access control. This transaction will be encapsulated in a new block and the ACC will not be executed until the new block is mined and included in the blockchain by some miner.
  • Step 4: During the access control process, the ACC will send a message to call the JC, if some potential misbehavior of the subject is detected.
  • Step 5: Once the JC  judges the misbehavior and determines the penalty, it will return the penalty to the ACC.
  • Step 6: Finally, the access result will be returned to both the subject and object, after the access control process finishes.

To demonstrate the feasibility of the framework, we provide a case study in an IoT system with one desktop computer, one laptop and two Raspberry Pi single-board computers, as shown in the figure below.

And the results of access control at both the subject and object sides are illustrated in the figures below.

光メディアインタフェース研究室 2018

光メディアインタフェース研究室 助教の田中です.前回の投稿は,2015年3月「光メディアインタフェース研究室 本格始動!」でした.あの時は,研究室が発足したばっかりでしたが,いまや学生はM1からD3まで勢揃いの立派な研究室になりました.今回は,これまでの本研究室の成果を振り返ってみます.

光メディアインタフェース研究室は,1.機械がカメラを通して現実を理解する「コンピュータビジョン」 2.あらゆる物体の質感をリアルに再現する「コンピュータグラフィクス」 3.カメラの常識を打ち破る新しい撮影技術「コンピューテーショナルフォトグラフィ」 4.あらゆる研究の基礎となる「光学設計」の4本柱で研究を進めており,人間と機械が光を媒体としてシーンに関する情報を共有できる新しいインタフェースの実現を目指し活動しています.詳しい研究内容は,研究室ホームページ  をご参照ください.


今,光メディアインタフェース研究室には,教授・准教授・2名の助教の計4名の教員,事務補佐員,6名の博士後期課程学生,16名の博士前期課程学生,さらに1名のインターン生で活動しています.研究室の学生は,毎年何らかの研究予算を自分で獲得してきてプロジェクトを進めるなど,活発な人が多く在籍しています.CICPと呼ばれる学内予算はもとより,日本学術振興会特別研究員や,JST ACT-I,IPA未踏スーパークリエーター認定などを得ており,自ら未開拓分野を切り拓いていく姿勢には感心します.これまでに,本研究室を修了した学生は,3年間で延べ21人となり,各方面で活躍しています.「光メディアインタフェース研究室出身はすごい」と呼ばれるよう,より一層の活躍を期待したいところです.また,これまで多数の留学生・インターン等を受け入れています.なぜだか分かりませんが,多くがフランス語圏からの留学生です.もしかして,フランス語の案内ページも作ったほうが良いのでしょうか…?

本研究室は,活発に外部の機関と共同研究を実施してきました.複数の国内大手電機メーカーや農業・医療・自動車産業,アニメ産業の企業と連携中あるいは連携してきました.アカデミックでは,東京大学・京都大学・大阪大学・九州大学・国立情報学研究所・筑波大学・広島大学・早稲田大学・東邦大学や,Carnegie Mellon University (CMU・アメリカ)・Arizona State University(アメリカ)・The University of Picardie Joule Verne (フランス)・The University of California Los Angeles (UCLA・アメリカ) の研究者と共同研究を実施してきました.プロジェクトの数が多く,学生が多数の幅広いテーマから自分のやりたいことを選択し取り組めるのも,この研究室の大きな特徴かもしれません.

彼らとともに成し遂げた研究プロジェクトは から見ることができます.数字で見ていくと,情報電子通信学会論文賞やMIRU長尾賞などの大きな賞を含む24個の受賞,2本のPAMI(最難関国際論文誌)を含む12本の論文誌および5本のCVPR(最難関国際会議)を含む21本の査読付き国際会議論文を発表してきました.どのテーマで研究しても,在学中に1回以上の発表を学会で行うべく,ハイレベルな研究活動を行っています.





2つ目の研究は,MIRU長尾賞という,いわゆる日本一と認められた,新しいコンピュータビジョンの研究です.コンピュータビジョンというと,カメラで撮影された視覚情報をコンピュータでどのように理解するか,という風に説明されてきました.我々の研究室では,カメラで撮影された後の2次元画像からスタートしているようでは情報が不十分であり,画像になる前の,現実世界を飛び交う光の状態からシーン理解を行うべきだという信念のもと,より次元を増やした計測からシーン理解を行う研究を行っています.その中でも,時間の次元を増やす,ということに関して,世界のトップレベルとしのぎを削っています.時間の次元と言っても,単にビデオやハイスピードカメラで撮影するということではありません.この研究室で行っているのは,「光」自体が伝わっていく様子を可視化できるくらいの超高速な計測,時間で言うと数ナノ秒(0.000000001秒)から数ピコ秒(0.000000000001秒)です.図2に,実際に 250ピコ秒で撮影した画像を載せています.このレベルになると,光がどのように反射・散乱し,シーン中を伝わっていくのかが分かります.当然ながら,こうした情報を使うと,3次元形状,物体の材質,異常検知などが簡単に行えるようになります.




最後に,NAISTへの入学を検討している方や,インターン希望の方,ぜひ一度光メディアインタフェース研究室に見学に来てください.文章ではわからない研究のこと,研究室の雰囲気,研究の進め方,などなど,今後の進路に役立つこと間違いありません.NAISTでは,受験生のためのオープンキャンパス を毎年2月と5月に行っています.また,受験を検討している人のために2・3日の研究体験ができるサマーセミナー (8月頃)や スプリングセミナー (2月頃)もやっています.さらに,いつでも見学会 という制度を使うと,研究室が受け入れ可能な日はいつでも見学・相談が可能です.ぜひ,ご検討ください.

Speed Up Machine Learning Through Efforts of Advanced VLSI Circuits

Hello, world. This is from Assistant Professor Renyuan Zhang with Computing Architecture Lab. to share some progresses from our group, which are all about speeding up the machine learnings through efforts of advanced VLSI circuits.

Why do we implement the machine learning in silicon?

The artificial intelligence plays very important roles in the modern/post IoT. A common challenge is how to efficiently implement machine learning algorithms in systems with the cloud-edge prototype. Along with the development of computing processors, one option is to distribute the machine learning at the “edge” of systems as show in the sided figure.

In such a system, both of learning and recognition (for instance) are carried out at the edge instead of central stations, which is helpful to reduce the communications and able to active movable equipment off-connecting any PC (such as vehicles and body-area). Unfortunately, very few on-chip learning processors have been developed for machine learning at edge in silicon.

Both of digital and analog efforts have been done a lot to develop on-chip learning processors. Multi-computational-core is a general digital strategy to process a large number of complex computations in parallel (see machine learning works by GPGPU). Our group members also hold some experiences of implementing machine learning by CGRAs with very high parallelism. For many IoT tasks, the computational accuracy is not extremely required. Therefore, some approximate computing processors were designed by analog circuits. In some of my early works, it is found the chaos of analog signals greatly speeds up the learning process. We have realized ultra-high speed on-chip learning for some specific algorithms such as SVM or K-means. In summary, the CPU or GPGPU is not always applicable for every edge device; analog processors have potentials of high-speed and low-power, but their functions (target algorithms) are always fixed and specific. Therefore, most of current/previous IoT works, which employ machine learning, have to carry out the “learning” centrally, then do the recognitions locally.

What have we done?

On the academia side, the ML and AI research/development is one of the most popular fields. There are many groups and societies are trying to design specific VLSI systems for realizing efficient ML. Most of those works are based on the high performance digital processors. The analog (even physics device) circuit has been a new (or we can say re-born) trend leading to ultra-high speed and efficiency. Our early work was one of world-first pure-analog ML chips with benefits but suffered from poor generalization. Thus, we are going to develop generally and practically feasible technologies in this field. In industry, many software/hardware makers start their competitions on VLSI chips for ML/AI in past five years (see “IBM Scientists Show Blueprints for Brain-like Computing,” “Building a Brain on a Silicon Chip” and “Intel Reveals Neuromorphic Chip Design”). The mobile supplier HUAWEI even actually applied the NPU (Neural Processing Unit) in their newest mobile chip “Kirin970”.  In our early works, world fastest learning (64-D SVM learning within 0.1us, 2012; 64-D K-means within 20ns, 2013; 64-D SVDD within 40ns, 2014) has been achieved (see the sided figure as an example: SVDD algorithm is implemented by VLSI to do the multi-class image classification). However, the algorithms and capacities were fixed. We expect to keep the speed benefit and generalize the target algorithms.

What are we going to do?

As an exploration, we are expecting to develop general purpose on-chip learning processor which accelerates various machine learning processes by analog VLSIs. As consequence, local/on-chip machine learning is feasible and practical for ubiquitous applications.

We are designing ultra-high speed and low-energy hardware (VLSIs) for implementing the so-called on-chip learning (not for specific but various algorithms). The analog approximate computing strategy is adapted. The circuit-, unit-, and architecture-level are explored; several actual VLSI chips are expected to fabricate as demonstrations. For instance, we will demo the visual tracking with ML by a single VLSI chip without any PC.

As shown in the sided figure, three key ideas are included in this project: 1. General purpose Analog Calculation Unit (called ACU); 2. Fully-parallel accelerating strategy; 3. Expanding the learning capacity. It is a cross-field among high-performance-computing, the VLSI designs and machine learning theories.

What are we expecting?

Firstly, we expect our novel learning mode: chaos of continuous signals, to impact the machine learning societies for reasonable applications. Along with the development of IoT, various demands of on-chip machine learning require the well-performance hardware besides CPU or GPGPU. Secondly, a general purpose ML accelerator leads to practical “smart chips”, which are embedded into the edge-devices of IoT. Currently, most of IoT works execute the recognition at the edge, but do the learning process on cloud or central stations. The edge learning offers a potential for industry to build more efficient and intelligent network, in which the machine learning can be done by VLSI chips instead of CPUs.

Most importantly, the road-map of VLSI scaling-down might reach the end soon. Different strategies of high performance computing architecture should be explored. The progress of this project is expected to push the analog approximate computing fashion into practical fields. The developed ACUs can be applied in various architectures even together with digital frameworks. The general purpose and programmable analog computing, as an internal progress, offers a different option of data processing: lighter, but faster.

Thank you very much for your kind reading.






表1: 電気特性と熱特性の類似性

電気特性 温度特性
電圧 (V) 温度 (°C)
抵抗 (Ω) 熱抵抗 (°C/W)
容量 (F) 熱容量 (J/°Cm3)
電流 (A) 熱量 (J)
消費電力 (W) 熱流量 (J/s)



図1: 電気特性モデルと熱特性モデル


  1. 温度を考慮した電気特性モデルパラメータの推定
  2. 熱特性モデルパラメータの推定




図2: 温度毎の電流特性とそのモデル。点が電流の実測値で、線がシミュレーションによる計算結果を表す。



図3: フィードバック制御図


図4: パルス幅変調回路による電力保持機構



図5: バックコンバータの回路図とプリント基板への実装


図6: 電気特性(左)と熱特性(右)の過渡変化の実測とシミュレーションの比較






ロボットというのは往々にして生物を手本にその機構や制御を設計します.何故なら,生物というのはその誕生から進化をし続けて現在の環境へと最適化されているからです.そんな生物の機構や制御を紐解けば,ロボットにとっても最適な機構や制御を導くことができます.例えば私の過去の研究では,ヒューマノイドロボットの歩行制御を人の歩行の特徴,モデルを基に設計することで人のような省エネな歩行を実現してきました.最近では,生物の運動制御を司る小脳の神経回路網を模倣したReservoir Computingを利用した継続学習に挑戦しています.



では,そんな複数の移動形態を扱うにはどうすれば良いのでしょうか?単純に考えれば,各移動形態を実現する制御器を別々に用意すれば良いと思うでしょう.しかしこの移動形態というのは細かく分ければ無数に存在しますので,全てを別の制御器で扱うというのは現実的ではありません.そのため,複数の移動形態を可能な限り統一的に扱うような制御器を設計することが求められます.そんな制御器として,環境との受動的な点接触と全身関節角が接触角に連動して動く仮想拘束の2つの原理を組み込んだPassive Dynamic Autonomous Control (PDAC)という手法を提案してきました.この2つはロボットの全身ダイナミクスを非常に簡単なものにしてくれ,さらには,環境と何処で接触しているのか,どのように連動させるか,の2点だけを考えれば全ての移動形態を制御できるようにしてくれます.



生物の何処が身体を上手に制御しているかと言えば,それは間違いなく脳だと言えるでしょう.つまり,生物の脳構造を紐解いて模倣することができれば,生物のように巧みな動きが実現できるはずです.脳構造を模倣したもの全般はニューラルネットワークと呼ばれ,最近ではこのネットワークの層構造を何十,何百層とした,いわゆる深層学習が認識技術として大きな成功を収めています.この深層学習は人の視覚野と似た情報処理構造をしているとされていますが,では運動制御を司っている小脳を模倣した情報処理構造はどうなっているのでしょうか?その答えがLiquid State Machine (LSM)やEcho State Network (ESN)に代表されるReservoir Computingになります.

Reservoir Computingは入力層・リザーバ層・出力層の3層で構成され,このリザーバ層が自身の信号を再び取り入れるという,再帰的な結合を持つリカレントニューラルネットワークの一種となります.Reservoir Computingの最たる特徴は,学習するノード間の結合がリザーバ層と出力層を接続しているリードアウトのみ,という点にあります.つまり,一般的なニューラルネットワークと異なり,出力層における誤差を逆伝播させることなく,リザーバ層は初期のランダムな結合のままで固有のダイナミクスを有したままとなります.このリザーバ層の構造が小脳における顆粒層に相当するといわれています.また,リザーバ層を学習しないで良いため,そのノード数は通常のニューラルネットワークより非常に大きくしやすくなっており(1千〜1万個など),顆粒層(1000億個ほど)と比べればまだまだ少ないですが,それでも類似した特徴といえます.

最近は,このReservoir Computingを用いた継続学習について研究しています.一般的なニューラルネットワークというのは,誤差逆伝播によってネットワーク全体の結合荷重を更新して学習したいタスクを獲得するため,タスクAについて学習済みのネットワークでタスクBを学習すると,タスクAを扱えなくなる「破滅的忘却」という問題を抱えています.Reservoir Computingであれば,リードアウトをスパースに学習するだけでタスク間の競合を最小限に抑えられます.また,リザーバ層にあるタスク入力を加えれば,固有のダイナミクスに基づいた入力依存の活性化パターンが期待されますので,タスクごとに異なるノードが活性化しやすいようになります.こういった工夫を取り入れることで従来のように全てのタスクについて同時に学習するのではなく,生物のようにタスクを1つずつ継続的に学習することができます.この継続学習は,生物のように生涯を通じて学習し続けて次々に新しいことができるようになる自律ロボットを開発するにあたって非常に重要な能力になります.