Biometrics-based Recognition System on Wearable Devices Weitao Xu
Biometrics-based Recognition System on Wearable Devices Weitao Xu University of Queensland Supervisor: Prof. Neil Bergmann Dr. Wen Hu Outline
Introduction Project 1-Face recognition Project 2-Key generation Project 3-Gait recognition Conclusion 2 of 51 Introduction What is biometrics? The users behavioral or physiological characteristics to determine or verify an identity.
Fingerprint Handwritten signature Face 3 of 51 Introduction Wearable devices Smart watch
Smart glass Smart wristband 4 of 51 Introduction Sensors Accelerometer Gyroscope Compass Camera Micphone
GPS Bluetooth Applications Navigation Localization
Health monitoring Tracking Remote control 5 of 51 Research projects Biometrics 1 2
Wearable device 3 face recognition on smartglass Key generation system for wearable devices Gait recognition system using energy harvester
6 of 51 Research project 1 Sensor-assisted face recognition system on smartglass 7 of 51 Background Face recognition Security
Friends Tagging Face Unlock Smart glasses: a third eye 8 of 51 Main Challenges Short battery life Challenges
Limited processing capability Depend on cloud architecture Possible Solution Wireless transmission is expensive Our goal Robust and efficient in-situ face recognition system on smart
glass 9 of 51 System Architecture Step 1: capture multi-view face images Step 2: sampling optimization MASO Step 3: face recognition Target MVSRC
User ID User MASO: Max Accuracy Sampling Optimization MVSRC: Multi-view Sparse Representation Classification 10 of 51 MVSRC MVSRC exploits the sparsity of face images from different
views and apply a weighted model to improve accuracy. 11 of 51 Max Accuracy Sampling Optimization Problem: applying MVSRC directly on a face image sequence is expensive Observation: adjacent frames are similar Solution: downsample a face image sequence What is a good down sampling strategy? 12 of 51
Max Accuracy Sampling Optimization We propose an optimization strategy based on machine learning. We relate the face recognition accuracy with angles of face images We use IMU sensors(accelerometer, gyroscope, magnetometer) to estimate the angles of the face images We use support vector regression (SVR) to downsample face images. Recognition process 13 of 51
Obtain Angle Information Target Final view angle final First view angle start User start .
1 User 2 k final 14 of 51 Evaluation Datasets
Honda/UCSD dataset: 20 subjects 59 image sequences Frames of each video clip vary from 12 to 645 Private dataset: 10 subjects (2 females and 8 males) 9 different combinations of expressions and environments 10.8K face images 15 of 51 Evaluation of MVSRC
Comparison with competing multi-view face recognition methods Private dataset Honda dataset MVSRC is 5%-10% more accurate than other methods 16 of 51 User study
Implementation: Vuzix M100 Smart glass 5 users, 10 subjects Different environments Comparison with OpenCV face recognition methods 7%-15% Statistic Our system Eigenfaces Computation
time 516-582 ms 247-331 ms Energy consumption 506-535 mJ 316-410 mJ Expected
battery life 0.28 hr 0.37 hr Memory Usage 55-64 MB 38-44 MB Same level
17 of 51 Demo https://www.youtube.com/watch?v=lVRS4e3Glho 18 of 51 Research project 2 Key generation system for wearable devices 19 of 51 Background
Wearable devices smartphone, smart glass, smart watch etc. Wearable devices pairing Data exchange 20 of 51 Traditional Methods
Explicit Input Labor-intensive Small form factor 21 of 51 Traditional Methods Key Exchange Public key infrastructure Cannot distinguish legitimate (D-H protocol) devices and attackers
22 of 51 Traditional Methods Static key Vulnerable to attackers Key revocation 23 of 51 System Overview Walkie-Talkie: On-body Authentication
Automatic Secure Pairing Spontaneous Key Generation Devices on the same body measure the same gait signals 24 of 51 Model User model: only the devices on the same body can be paired together Adversarial model: Active imposter Passive imposter
25 of 51 Design details Signal Processing Alice REQ Key generation Secure Communication
RSP Bob Source Separation REQ Signal Alignment RSP
Data Collection Quantization Signal Processing Reconciliation Key Generation Privacy
Amplification Data Sharing 26 of 51 Signal processing- signal alignment Space alignment -Coordinate system of different system are not the same -We transform the acceleration signal to body reference system 27 of 51
Key Generation Gravity direction Acc Quantization Sideway direction Reconciliation Walking direction Privacy Amplification
LAlice=[1,3,5,6,7,9,10] LBob=[1,2,5,6,8,9,10] Reserve common indexes L=[1,5,6,9,10] Reserve key generated at same index KAlice=10101 KAlice=1101001 KBob=10101 KBob=1001101 30 of 51 Key Generation-privacy amplification Privacy amplification: Correlation exits as walking is repetitive
Reconciliation reveal information to attackers Steps: Segmentation K1Alice(30bits) K2Alice .. KAlice=11010011011001011. Perform XOR KAlice=[(K1Alice XOR K2Alice) ,(K3Alice XOR K4Alice),.]
31 of 51 Evaluation Dataset 20 subjects (14 males, 6 females) 32 of 51 Evaluation Impact of reconciliation
Reconciliation reduces bit rate, however it can improve bit agreement rate up to 100%. 33 of 51 Evaluation Security analysis Passive imposter Active imposter Mutual Information among different devices
An active imposter can only achieve approximately 50% bit agreement rate (=0.8)) 34 of 51 System Implementation Implementation Moto E2 smartphone Bluetooth Low Energy (BLE) mode AES encryption/decryption Computation Time (ms)
Energy Consumption (mJ) Signal processing 108.3 71.2 Key generation 208.1 12.7
AES encryption 0.2 0.1 AES decryption 0.2 0.1 Total
316.8 85.6 Account to 0.002 of the battery supply only 35 of 51 Demo https://www.youtube.com/watch?v=YBFBJrNZy48 36 of 51
Research project 3 Gait recognition system for wearable devices 37 of 51 Background-gait recognition Gait recognition refers to identify an individual by his/her unique walking pattern. Mission: Impossible 38 of 51
Background-energy harvester Kinetic energy harvester (KEH): - generate power from kinetic motions like walking, running 39 of 51 Background-energy harvester AMPY-world's first wearable motion-charger SolePower-self-sustaining smart boot 40 of 51
Motivations Accelerometer-based gait recognition - continuous sampling consumes a lot of energy Our goal: - use generated voltage as source signal to realize gait recognition 41 of 51
System overview Traditional accelerometer-based gait recognition system: Proposed KEH-based gait recognition system: Accelerometer Save energy 42 of 51 Prototype design Piezoelectric energy harvester (PEH) Electromagnetic energy harvester (EEH)
43 of 51 Data collection Dataset 20 subjects (14 males, 6 females) Indoor, Outdoor 44 of 51 Feasibility study 45 of 51 Evaluation results
Comparison with accelerometer-based system KEH-Gait can achieve comparable recognition accuracy when we use more than 5 gait cycles. 46 of 51 Energy consumption analysis Our system can reduce energy consumption by 75.8)8)%. 47 of 51 Demo
https://www.youtube.com/watch?v=LOqSq28oexA 48 of 51 Conclusion We implement three biometrics-based recognition system on wearable devices We conduct extensive evaluations to evaluate their performance and security.
We implement the systems on commercial devices to demonstrate the feasibility. With the ubiquity of wearable devices, more security and privacy issues will arise. Future work will focus on addressing these challenges on emerging wearable devices. 49 of 51
References 1. Weitao Xu, Yiran Shen, Neil Bergmann,Wen Hu. Sensor-assisted Face Recognition System on Smart Glass via Multi-view Sparse Representation Classification IPSN 2016 (CORE A*) 2. Weitao Xu, Girish Revadigar, Chengwen Luo, Neil Bergmann,Wen Hu.Walkie-Talkie: Motion-Assisted Automatic Key Generation for Secure On-Body Device Communication IPSN 2016 Best paper runner-up award (CORE A*) 3. Weitao Xu, Guohao Lan, Qi Lin, Sara Khalifa, Neil Bergmann, Mahbub Hassan, Wen Hu. KEH-Gait: Towards a Mobile Healthcare User Authentication System by Kinetic Energy Harvesting conditionally accepted by NDSS' 2017 (CORE A) 50 of 51
Thanks for your attention! Questions and suggestions? 51 of 51
Illness Anxiety disorder. Formerly called Hypocondriasis. Illness Anxiety Disorder is a somatic symptom disorder characterized by the misinterpretation of normal bodily functions as signs of serious illness. People with this disorder fear or mistakenly believe that normal bodily reactions represent...
and the fact that many cephalopods were nektonic . make them excellent guide fossils. Mollusks. The Ammonoidea, cephalopods with wrinkled sutures, constitute three groups: the goniatites, ceratites, and ammonites. Ammonites, while present during the entire Mesozoic,
Tropic hormones control the function Trophic hormones promote survival and growth of targets Releasing Hormone Release-inhibiting Hormone Hypothalamo-hypophyseal portal system A A S S P P P P P P P P P P P S 1 Tropic hormones control...
Although evidence for non-pharmacological interventions was not reviewed, harms of such approaches are minimal ... Parenteral availability of haloperidol may still warrant short-term use in emergencies. ... excretion, interactions with other medications longer time to reach steady state levels ...
Kinematics in 1-D. Learning Target. I can differentiate between position, distance, displacement, speed, and velocity. What is motion? The mathematical description of motion is called kinematics. The explanation of motion in terms of its causes is called .
Unquestionably monotheistic—this is why Judaism, despite its size, is so important to world religions ... This begins the notion of prophecy which is central to all Abrahamic religions. ... illustrated values of Judaism, and importance to the Jewish faith—think: if...
* VI. Proton, Neutron and Electron Amu = 1/12 the mass of a carbon atom containing 6 protons and 6 neutrons. This makes the proton ~1amu. The proton and electron have electrical charge * A. Elements What makes elements unique...
Ready to download the document? Go ahead and hit continue!