![Page 1: Anomaly Detection in X-ray security imaging · in X-ray security imaging. Lewis Griffin. Computer Science, University College London, UK. ... - X-Ray Operator. The information in](https://reader033.vdocuments.us/reader033/viewer/2022042417/5f32dce232d17977840e5a62/html5/thumbnails/1.jpg)
Anomaly Detectionin X-ray security imaging
Lewis GriffinComputer Science, University College London, UK
![Page 2: Anomaly Detection in X-ray security imaging · in X-ray security imaging. Lewis Griffin. Computer Science, University College London, UK. ... - X-Ray Operator. The information in](https://reader033.vdocuments.us/reader033/viewer/2022042417/5f32dce232d17977840e5a62/html5/thumbnails/2.jpg)
![Page 3: Anomaly Detection in X-ray security imaging · in X-ray security imaging. Lewis Griffin. Computer Science, University College London, UK. ... - X-Ray Operator. The information in](https://reader033.vdocuments.us/reader033/viewer/2022042417/5f32dce232d17977840e5a62/html5/thumbnails/3.jpg)
A vision for automated security X-ray analysis
illegal dangerous valuable unusual
dangerous valuablelegal
shielding semantics anomaly
A S AgS
![Page 4: Anomaly Detection in X-ray security imaging · in X-ray security imaging. Lewis Griffin. Computer Science, University College London, UK. ... - X-Ray Operator. The information in](https://reader033.vdocuments.us/reader033/viewer/2022042417/5f32dce232d17977840e5a62/html5/thumbnails/4.jpg)
What does the AI revolution offerfor Semantic Analysis in X-ray security?
The Promise
better than human performance!
The Problem
massive training datasets needed
Summary: Deep Learning methods are powerful enough, the main problem is logistical –assembling massive labelled datasets.
![Page 5: Anomaly Detection in X-ray security imaging · in X-ray security imaging. Lewis Griffin. Computer Science, University College London, UK. ... - X-Ray Operator. The information in](https://reader033.vdocuments.us/reader033/viewer/2022042417/5f32dce232d17977840e5a62/html5/thumbnails/5.jpg)
illegal dangerous valuable unusual
dangerous valuablelegal
shielding semantics anomaly
A S AgS
A-R = RelativeWhy does that pallet of lemons look different from that one?
P-R = Passenger-relativeStrange luggage for a business traveller.
F-R = Flight-relative…
‘Unexpected item in the bagging area’:Anomaly Detection in X-ray Security Images
Lewis D. Griffin, Matthew Caldwell, Jerone T. A. Andrews an Helene Bohler
![Page 6: Anomaly Detection in X-ray security imaging · in X-ray security imaging. Lewis Griffin. Computer Science, University College London, UK. ... - X-Ray Operator. The information in](https://reader033.vdocuments.us/reader033/viewer/2022042417/5f32dce232d17977840e5a62/html5/thumbnails/6.jpg)
stream-of-commerce(n=5000)
staged threat(n=234)
3% 15% 81%
unrecognizable
increasing threat visibility
easily visibledifficult to see
A Parcel Dataset
Threat Detection for this problem is easy. With100 threat examples, we can build a system thatdetects 99% of threats for 1% false alarms.
AUC
number of training SoC images =number of training staged-threat images
![Page 7: Anomaly Detection in X-ray security imaging · in X-ray security imaging. Lewis Griffin. Computer Science, University College London, UK. ... - X-Ray Operator. The information in](https://reader033.vdocuments.us/reader033/viewer/2022042417/5f32dce232d17977840e5a62/html5/thumbnails/7.jpg)
Representation is the Keytraining testing
supervised
Trained on male/female label
anomaly detection
Trained on same-person pairs vs. different person pairs
![Page 8: Anomaly Detection in X-ray security imaging · in X-ray security imaging. Lewis Griffin. Computer Science, University College London, UK. ... - X-Ray Operator. The information in](https://reader033.vdocuments.us/reader033/viewer/2022042417/5f32dce232d17977840e5a62/html5/thumbnails/8.jpg)
Image Patch
multiple convolution,pooling and rectification layers
pooling layer
logit layer
lant
ern
car
… … …
softmax layer
semanticconfidences
threshold
appearancerepresentation
semanticrepresentation
Convolutional Neural Network (CNN) trained to recognize objects in photographs
Representations used for Anomaly Detection
![Page 9: Anomaly Detection in X-ray security imaging · in X-ray security imaging. Lewis Griffin. Computer Science, University College London, UK. ... - X-Ray Operator. The information in](https://reader033.vdocuments.us/reader033/viewer/2022042417/5f32dce232d17977840e5a62/html5/thumbnails/9.jpg)
Schematic of Automated Anomaly Detection
stream-of-commerce parcels
distribution of SoCappearance
01
01
appearance-representationfrom anobject-categorizing CNN
encode
test image
01
01
outlier?
![Page 10: Anomaly Detection in X-ray security imaging · in X-ray security imaging. Lewis Griffin. Computer Science, University College London, UK. ... - X-Ray Operator. The information in](https://reader033.vdocuments.us/reader033/viewer/2022042417/5f32dce232d17977840e5a62/html5/thumbnails/10.jpg)
Anomaly Detection Results. 1
anomaly
pass
SoC
staged threats
![Page 11: Anomaly Detection in X-ray security imaging · in X-ray security imaging. Lewis Griffin. Computer Science, University College London, UK. ... - X-Ray Operator. The information in](https://reader033.vdocuments.us/reader033/viewer/2022042417/5f32dce232d17977840e5a62/html5/thumbnails/11.jpg)
AUC
appearance
semantic
appearance + semantic
number of SoC images1 8 64 512 4096
• Peak performance requires many more images than for threat detection.• But only unlabelled SoC images are required.• Semantic less effective than appearance.• Small boost from combination
Anomaly Detection Results. 2
![Page 12: Anomaly Detection in X-ray security imaging · in X-ray security imaging. Lewis Griffin. Computer Science, University College London, UK. ... - X-Ray Operator. The information in](https://reader033.vdocuments.us/reader033/viewer/2022042417/5f32dce232d17977840e5a62/html5/thumbnails/12.jpg)
Anomaly Detection Results. 3
falsealarm
beststandard approach
ourapproach
18% false alarms 65% false alarms
90% detection
50% catch
×15 discovery ×3 discovery
![Page 13: Anomaly Detection in X-ray security imaging · in X-ray security imaging. Lewis Griffin. Computer Science, University College London, UK. ... - X-Ray Operator. The information in](https://reader033.vdocuments.us/reader033/viewer/2022042417/5f32dce232d17977840e5a62/html5/thumbnails/13.jpg)
Anomaly Detection works.Possibly well enough to havea role, but improvedperformance is desirable.
Summary
Questions?
Semantics: Its all aboutlabelled datasets
![Page 14: Anomaly Detection in X-ray security imaging · in X-ray security imaging. Lewis Griffin. Computer Science, University College London, UK. ... - X-Ray Operator. The information in](https://reader033.vdocuments.us/reader033/viewer/2022042417/5f32dce232d17977840e5a62/html5/thumbnails/14.jpg)
Innovation Workshop Partnered with ICAO Working Group on Innovation in Aviation Security (WGIAS)
Bruno Faviero, Co-Founder and COO, Synapse Technology
![Page 15: Anomaly Detection in X-ray security imaging · in X-ray security imaging. Lewis Griffin. Computer Science, University College London, UK. ... - X-Ray Operator. The information in](https://reader033.vdocuments.us/reader033/viewer/2022042417/5f32dce232d17977840e5a62/html5/thumbnails/15.jpg)
Bruno FavieroCo-Founder and COO
The information in this document should be treated as proprietary and confidential to Synapse Technology Corporation
AVSEC World: ICAO Innovation Workshop February 28, 2019
![Page 16: Anomaly Detection in X-ray security imaging · in X-ray security imaging. Lewis Griffin. Computer Science, University College London, UK. ... - X-Ray Operator. The information in](https://reader033.vdocuments.us/reader033/viewer/2022042417/5f32dce232d17977840e5a62/html5/thumbnails/16.jpg)
The information in this document should be treated as proprietary and confidential to Synapse Technology Corporation
Silicon ValleyCenter of Innovation
![Page 17: Anomaly Detection in X-ray security imaging · in X-ray security imaging. Lewis Griffin. Computer Science, University College London, UK. ... - X-Ray Operator. The information in](https://reader033.vdocuments.us/reader033/viewer/2022042417/5f32dce232d17977840e5a62/html5/thumbnails/17.jpg)
The information in this document should be treated as proprietary and confidential to Synapse Technology Corporation
“If I had asked people what they wanted, they would have said faster
horses"
Henry Ford
![Page 18: Anomaly Detection in X-ray security imaging · in X-ray security imaging. Lewis Griffin. Computer Science, University College London, UK. ... - X-Ray Operator. The information in](https://reader033.vdocuments.us/reader033/viewer/2022042417/5f32dce232d17977840e5a62/html5/thumbnails/18.jpg)
The information in this document should be treated as proprietary and confidential to Synapse Technology Corporation
horses
Henry Ford
In the face of unknowable future threats, prioritize solutions that deliver speed and flexibility for maximum responsiveness.
![Page 19: Anomaly Detection in X-ray security imaging · in X-ray security imaging. Lewis Griffin. Computer Science, University College London, UK. ... - X-Ray Operator. The information in](https://reader033.vdocuments.us/reader033/viewer/2022042417/5f32dce232d17977840e5a62/html5/thumbnails/19.jpg)
The information in this document should be treated as proprietary and confidential to Synapse Technology Corporation
Humans are Human
• Long shifts
• Repetitive task
• Motivation
• Fatigue
![Page 20: Anomaly Detection in X-ray security imaging · in X-ray security imaging. Lewis Griffin. Computer Science, University College London, UK. ... - X-Ray Operator. The information in](https://reader033.vdocuments.us/reader033/viewer/2022042417/5f32dce232d17977840e5a62/html5/thumbnails/20.jpg)
The information in this document should be treated as proprietary and confidential to Synapse Technology Corporation
Lots of data generated
Highly repetitive
Specific list of items
Few seconds to make decision
AI is a perfect fit for checkpoint security
“Speed matters. Threats emerge quickly. And they evolve fast”.
Time
Performance
Machine
Human
![Page 21: Anomaly Detection in X-ray security imaging · in X-ray security imaging. Lewis Griffin. Computer Science, University College London, UK. ... - X-Ray Operator. The information in](https://reader033.vdocuments.us/reader033/viewer/2022042417/5f32dce232d17977840e5a62/html5/thumbnails/21.jpg)
The information in this document should be treated as proprietary and confidential to Synapse Technology Corporation
SHARPS
Syntech ONE™ pocketknife detection
Automatic Threat Detection
Syntech ONE™ handgun detection
FIREARMS
![Page 22: Anomaly Detection in X-ray security imaging · in X-ray security imaging. Lewis Griffin. Computer Science, University College London, UK. ... - X-Ray Operator. The information in](https://reader033.vdocuments.us/reader033/viewer/2022042417/5f32dce232d17977840e5a62/html5/thumbnails/22.jpg)
Dark Background
•Operator Assist = Operator + AI
•Operator usability and trust is important
•Deploy today
The information in this document should be treated as proprietary and confidential to Synapse Technology Corporation
Syntech ONE™
START
![Page 23: Anomaly Detection in X-ray security imaging · in X-ray security imaging. Lewis Griffin. Computer Science, University College London, UK. ... - X-Ray Operator. The information in](https://reader033.vdocuments.us/reader033/viewer/2022042417/5f32dce232d17977840e5a62/html5/thumbnails/23.jpg)
The information in this document should be treated as proprietary and confidential to Synapse Technology Corporation
Syntech ONE™ Traction
5+ Million Bags
NRT
ARN
PLS KIX
SJC
FPS
![Page 24: Anomaly Detection in X-ray security imaging · in X-ray security imaging. Lewis Griffin. Computer Science, University College London, UK. ... - X-Ray Operator. The information in](https://reader033.vdocuments.us/reader033/viewer/2022042417/5f32dce232d17977840e5a62/html5/thumbnails/24.jpg)
Dark Background
The information in this document should be treated as proprietary and confidential to Synapse Technology Corporation
Benefits of Platform
• Updates for new threats
• Detects threat components across lanes
• Learning from system-wide data
Functionality Beyond AI
![Page 25: Anomaly Detection in X-ray security imaging · in X-ray security imaging. Lewis Griffin. Computer Science, University College London, UK. ... - X-Ray Operator. The information in](https://reader033.vdocuments.us/reader033/viewer/2022042417/5f32dce232d17977840e5a62/html5/thumbnails/25.jpg)
Dark Background
The information in this document should be treated as proprietary and confidential to Synapse Technology Corporation
Benefits of Platform
• Intuitive UI
• Analytics
• Customized training
Human-Computer Interaction
“Synapse notices things that the operators don't — things that would have made it through the checkpoint without the system…
security has improved.” - X-Ray Supervisor
“I’m 40-60% more effective with Synapse.”
“I no longer want to quit my job.”
- X-Ray Operator
![Page 26: Anomaly Detection in X-ray security imaging · in X-ray security imaging. Lewis Griffin. Computer Science, University College London, UK. ... - X-Ray Operator. The information in](https://reader033.vdocuments.us/reader033/viewer/2022042417/5f32dce232d17977840e5a62/html5/thumbnails/26.jpg)
The information in this document should be treated as proprietary and confidential to Synapse Technology Corporation
Challenges to Innovation
Security hardware is closed
AI is held to a higher standard than humans
Regulatory framework needed for a system that improves in the field
![Page 27: Anomaly Detection in X-ray security imaging · in X-ray security imaging. Lewis Griffin. Computer Science, University College London, UK. ... - X-Ray Operator. The information in](https://reader033.vdocuments.us/reader033/viewer/2022042417/5f32dce232d17977840e5a62/html5/thumbnails/27.jpg)
The information in this document should be treated as proprietary and confidential to Synapse Technology Corporation
Screening Automation Roadmap
Similar to Hold Baggage Screening
On the path towards
• Operators will focus on resolving alarms, not looking at images
• Fewer operators needed to look at X-ray
• Detecting new threats or items that emerge
• Processing bags more quickly
• Taking advantage of new sensors and technologies
Syntech SOLUTION
• First of its kind to integrate the machine with AI
![Page 28: Anomaly Detection in X-ray security imaging · in X-ray security imaging. Lewis Griffin. Computer Science, University College London, UK. ... - X-Ray Operator. The information in](https://reader033.vdocuments.us/reader033/viewer/2022042417/5f32dce232d17977840e5a62/html5/thumbnails/28.jpg)
Dark Background
Innovation is not just about technology.It’s also about innovating on:
The information in this document should be treated as proprietary and confidential to Synapse Technology Corporation
PEOPLE PROCESS POLICY
“Coordination Innovation”How do we put it all together.
![Page 29: Anomaly Detection in X-ray security imaging · in X-ray security imaging. Lewis Griffin. Computer Science, University College London, UK. ... - X-Ray Operator. The information in](https://reader033.vdocuments.us/reader033/viewer/2022042417/5f32dce232d17977840e5a62/html5/thumbnails/29.jpg)
The information in this document should be treated as proprietary and confidential to Synapse Technology Corporation
www.gsin.group/resources
Tomorrow’s technology to prepare for tomorrow’s problems.
+