SLIDE 1
CyberPhysical systems for Security and Services
Antonio Rizzo, Alessandro Rossi, Francesco Montefoschi, Giovanni Burresi Carlo Festucci, Maurizio Caporali Siena, 14 Sett 2018
for Security and Services Antonio Rizzo, Alessandro Rossi, - - PowerPoint PPT Presentation
for Security and Services Antonio Rizzo, Alessandro Rossi, - - PowerPoint PPT Presentation
CyberPhysical systems for Security and Services Antonio Rizzo, Alessandro Rossi, Francesco Montefoschi, Giovanni Burresi Carlo Festucci, Maurizio Caporali Siena, 14 Sett 2018 Case of study: ATM-Sense TREND Rapine vs Attacchi ATM - Fonte
SLIDE 2
SLIDE 3
TREND Rapine vs Attacchi ATM - Fonte ABI 2017
200 400 600 800 1.000 1.200 1.400 2012 2013 2014 2015 2016
Numero Rapine per Anno
SLIDE 4
SLIDE 5
ATM attacks
SLIDE 6
New Attacks
https://www.europol.europa.eu/newsroom/news/27-arrested-in-successful-hit-against-atm-black-box-attacks
SLIDE 7
Video Surveillance Approach
SLIDE 8
Intel RealSense Depth Cameras
- Powerful Open Souce SDK
- Easily Embeddable
SLIDE 9
Intel RealSense Depth Cameras
SLIDE 10
Convolutional Neural Networks
source image convolutions maxpool convolutions maxpool fully connected
SLIDE 11
Image Convolutions
SLIDE 12
Image Convolutions
SLIDE 13
Image Convolutions
SLIDE 14
Convolutions
SLIDE 15
Max Pooling
source image convolutions maxpool
SLIDE 16
More layers...
source image convolutions maxpool convolutions maxpool
SLIDE 17
Visualizing Convolutional Layers
References:
- Lee, H., Grosse, R.,
Ranganath, R., & Ng, A. Y. (2009, June). Convolutional deep belief networks for scalable unsupervised learning
- f hierarchical representations.
In Proceedings of the 26th annual international conference on machine learning (pp. 609-616). ACM.
SLIDE 18
Convolutional Neural Networks
source image convolutions maxpool convolutions maxpool fully connected
SLIDE 19
CNN: ImageNet Classification Error
References:
- Russakovsky, Olga, et
- al. "Imagenet large
scale visual recognition challenge." International Journal of Computer Vision 115.3 (2015): 211-252
- Hardware
Architectures for Deep Neural Networks, ISCA Tutorial, MIT
SLIDE 20
Machine Learning Process
- 1. Get a dataset
- 2. Define the network
architecture
- 3. Train and Test the
model
SLIDE 21
- 1. Get a Dataset
SLIDE 22
SLIDE 23
SLIDE 24
- 2. Define the Network Architecture
SLIDE 25
- 3. Train and Test the Model
SLIDE 26
Results
Test Dataset Classification Accuracy Background 98.19% Withdrawal 97.05% Attack 98.32% Average 97.85%
Five Frames analysis:
- No false alarms
- No undetected attacks
- Attack detection time:
○ mean: 2.4 sec ○ max: 3.3 sec Single Frame analysis:
Model Running on SECO SBC–A80 with Intel Braswell CPU
○ mean: 0.5 sec
SLIDE 27