Original Problem:
“We seek to develop tools that
will optimize discovery and
investigation of adversary
communication trends on
social media, allowing ARCYBER
and others to more efficiently
respond and mitigate threats
posed by enemy messaging.”
Team Narrative Mind
Sponsor: US Army Cyber Command (ARCYBER)
100
Ten weeks of interviews later... Current Status:
We learned a lot about this this
space and the acquisition
process.
Show similar commercial
tools--hear feedback
The Journey
Week #
EmotionalState
0 1 2 3 4 5 6 7 8
9
Suggestions for MVP
switches each week
Hashtag co-
occurrence
framework
ARCYBER
visit, spec of
final MVP
Unclear response from
stakeholders.
How do we begin?
Our sponsor gave us a lot of freedom to explore.
Initial Mission Model Canvas (Week 1)
- Gnip/Twitter
- CrowdFlower,
Samasource, or
Mechanical Turk
- Pre-existing social
media service and micro-
labor aggregators
- Optimize workflow for social
media analysts
- Expedite categorization
of social media content.
- Use MechanicalTurk to
crowdsource
categorization of content.
- - Algorithmic
virality
predictor
ARCYBER wants to
derive “meaning”
Primary: Intelligence
analysts receive a better
platform.
- Help intelligence analysts: receive cleaner, pre-categorized data,
- Architecture that can
support massive
concurrent data
aggregation and
analysis. E.g.
Storm/Hadoop.
- Testing with analysts
- MechanicalTurk or crowdsourcing labor (microtasks)
- UI Development/Testing with CYBERCOM/ARCYBER analysts.
- Software Development
- Access to Twitter
firehose (Stanford
academic license)
- Individual Analysts
- ARCYBER
- Continued partnership
with crowdsourcing firms,
CrowdFlower,
Samasource, etc.
Beneficiaries
Mission AchievementMission Budget/Costs
Buy-In/Support
Deployment
Value
Proposition
Key Activities
Key Resources
Key Partners
- Categorize
SM posts by
content
Week 1
How do we begin?
We learned about a wide variety of problems.
Weeks 1-3
DoD/Gvt. social media presence is weak.
Account bans and multiple aliases across
different networks make IDs hard to track
Can’t monitor who views dark-web content.
Fail to understand which narratives are the
most salient.
No baseline for monitoring/aggregating
use of tech;
Language/culture experts aren’t able to
work at scale.
Scale of social media makes manual
efforts painful.
Significant data management overhead.
Can’t determine actual scale
Problems
Weeks 1-3
How did we respond?
We brainstormed a lot of MVP’s.
Weeks 1-3
Proposed MVPs
Group
A
Group B
Weeks 1-3
Information Overload
Information overload!
Week 4
Information Overload
How do we conceptualize and organize
the challenges our users face?
Week 4
What kinds of problems are there?
Global
Tweet-level
Awareness
Response
Mapping the Problem Space
Week 4
What kinds of problems are there?
Global
Tweet-level
Awareness
Response
Mapping the Problem Space
Week 4
What kinds of problems are there?
Global
Tweet-level
Awareness
Response
Mapping the Problem Space
Week 4
Mapping the Problem Space
Week 4
What kinds of problems are there?
Global
Tweet-level
Awareness
Response
Mapping the Problem Space
Awareness Response
Week 4
Mapping the Problem Space
Global
Tweet-level
Awareness Response
Week 4
Mapping the Problem Space
Global
Tweet-level
Awareness Response
IO Org Chart
No baseline for monitoring/aggregating
use of tech
Automatic Narrative Detection
Language/culture experts aren’t able
to work at scale
Important Event Predictor
Preempt real world events
Persistent ID-Alias tracker
Account bans and multiple aliases across
different networks make IDs hard to track
Site Scraper
Need more cached information
access
Expedited Content Categorization
Scale of social media makes manual efforts
painful
Bot Detector
Can’t determine actual scale of info.
Virality Predictor
Understand which narratives are the most salient
Company F
Persistent ID-Alias tracker
Existing Products
Global
Tweet-level
Awareness Response
IO Org Chart
No baseline for monitoring/aggregating
use of tech
Automatic Narrative Detection
Language/culture experts aren’t able
to work at scale
Company C
Virality Predictor
Understand which narratives are the most salient
Counter-Narrative Generator
DoD/Gvt. social media
presence is weak
Company E
Site Scraper
Need more cached information
access
Company D
Expedited Content Categorization
Scale of social media makes manual efforts
painful
Company B
Bot Detector
.
Company A
Important Event Predictor
Preempt real world events
Opportunities
Global
Tweet-level
Awareness Response
Adversary IO Org Chart
No baseline for monitoring/aggregating
adversary use of tech; cyber targeting
Automated Narrative Detection
Language/culture experts aren’t able
to work at scale
Open
Opportunities
Our Beneficiaries
How can we understand our users and
the environment in which they operate?
Week 5-6
MMC - Week 6
- Track how groups use
technology over time.
- Gnip/Twitter/Facebook
- CrowdFlower,
Samasource, or
Mechanical Turk
- Third-party access
platforms for social
media
-Data visualization
-Content analysis
platforms
Primary
ARCYBER
-Bg. General (decision
maker)
-MAJ/LTC/COL
(operational plan)
-Analysts/Operators
(actionable insights)
COCOMs
-General (decision
maker)
-MAJ/LTC/COL
-Analyst/Operator
Secondary
Political Campaigns
-Campaign managers
-Supporters
Consumer Brands
-CMO
-Public Relations Team
- Optimize workflow for
social media analysts.
-Deliver insights to
commanders about
online environment.
-Insights into
responses against
narratives
-Detect narratives
emerging in real time
-Early warning on
emerging brand issues ???????????
-Enable faster problem awareness to problem response times for
decision-makers across organizations- UI Development/Testing with ARCYBER analysts.
- Software Development
- Research aggregation
- Access to Twitter
firehose or API
- Local language
speaking crowdsourcing
staff.
- Accurate testing for
intercoder reliability
- ARCYBER: Bg. General,
LTC, Strategic Initiatives
Group, OTA, Purchasing
PMs, End-User operator)
-COCOMs: OTA, Operators,
Purchasing PMs
-Political Campaign:
Opposition research team,
???
-Private Sector: CMO, ???
Beneficiaries
Mission AchievementMission Budget/Costs
Buy-In/Support
Deployment
Value
Proposition
Key Activities
Key Resources
Key Partners
Week 6
Value Proposition Canvas
Products
& Services
Web/Desktop
Application
Act on reports and
plans generated
- Little understanding of
ground-level nuances -
Little understanding of
proliferating platforms
Customer
Jobs
Gains
Pains
Gain
Creators
Pain
Relievers
- Narrative detection and topic
categorization
- Global awareness
- Reduce uncertainty of
decision making
- Add methodological rigor
to ARCYBER’s operations
Awareness
ARCYBER - BG General (Decision Maker)
Bg. General
(decision maker)
❏ Quickly understand key thematic points of organization's use of social
media put out by intelligence briefs.
MAJ/LTC/COL
(operational planer)
❏ Determine what types themes are rising in popularity and better identify
type of response
Analysts/Operators
(actionable insights)
❏ New movements can be understood and tracked with less direct
cooperation of experts.
Mission Achievement
ARCYBER
Big Picture Success Analogy:
“Most COCOMs and IO shops spend their whole day looking for a needle in a haystack: a user, a
post, an IP address. For narrative-level awareness, we need a strategy that helps us divide the
haystack into a bunch of smaller haystacks that don’t all look like same damn pile of hay.”
Low-Fidelity MVP
Week 7
How can we incrementally improve
ARCYBER’s existing workflow?
Customer Quote
“Things are changing everyday. We need
something that can help us with our long-term
strategy, regardless of how many times we have to
adjust our execution.”
Week 7
Example Criteria
Co-Occurring Hashtags
Week 7
Research Dataset
Dataset Features:
● 600k Unique Tweets
● Spanning October 2015 to May 2016
● 200k Unique Hashtag Combinations
Procedure:
1. Merged records into frequency table of hashtag co-
occurrences.
2. Manually coded 1300 most frequent hashtag sets
3. Visualized volume of these hashtag sets over time as
related to big-picture “narrative”.
Key Findings:
● Process is reasonably scalable.
● Could be implemented quickly by
ARCYBER to supplement
workflow.
● Deciding on narrative categories is
difficult.
● Need to further condition input
tweets to specific groups.
● How does hashtagged traffic
compare to total traffic?
Week 7
Low-Fidelity MVP
Weeks 6-8
How would any of these MVPs ever
make it into the hands of our sponsors?
First Steps: OTA
Proving a prototype and informing a requirement.
Innovation Challenge Initial Award Testing Evaluation
~$5M
1. Industry Day
2. Reqs. synopsis
3. Submit white papers
4. Evaluate papers
5. Proposals selected.
6. Technical discussion.
Original
requirements
synopsis
modified.
1. ACT office has
personnel working with
testers from all ranks.
2. Army Cyber Battle Lab
involved for
integration/concepts.
Budget Adoption
1. Requirement written.
2. Companies apply.
3. If OTA company is
selected, may skip as
far as MS-B.
4. Follow FAR/JCIDS.
1. Problem
requirements are
changed iteratively.
2. Phase objectives
and timeline are
flexible.
Weeks 6-8
Traditional and
non-traditional
contractors must
partner.
MMC - Week 7
- Track how groups
propagate narratives with
co-occurring hashtags.
- Gnip/Twitter/Facebook
- CrowdFlower,
Samasource, or
Mechanical Turk
- Pre-existing social
media service and micro-
labor aggregators
- Third-party access
platforms for social
media
-Data visualization
-Content analysis
platforms (Sens.ai,
Leidos)
Primary
ARCYBER
-Bg. General (decision
maker)
-MAJ/LTC/COL
(operational plan)
-Analysts/Operators
(actionable insights)
COCOMs
-General (decision
maker)
-MAJ/LTC/COL
-Analyst/Operator
Secondary
Political Campaigns
-Campaign managers
-Supporters
Consumer Brands
-CMO
-Public Relations Team
OTA in parallel for
product
development
Create dual-use
demand w/ PR
and political
campaigns
- Optimize workflow for
social media analysts.
-Deliver insights to
commanders about
online environment.
-Insights into responses
against broadcasting
narratives
-Enable faster problem awareness to problem response times for
decision-makers across organizations
- MechanicalTurk or crowdsourcing labor (microtasks)
- UI Development/Testing with ARCYBER analysts.
- Software Development
- Research aggregation?
- Access to Twitter
firehose or API
- Local language
speaking crowdsourcing
staff.
- Accurate testing for
intercoder reliability
- ARCYBER: Bg. General,
LTC, Strategic Initiatives
Group, OTA, Purchasing PMs,
End-User operator)
-COCOMs: OTA, Operators,
Purchasing PMs
-Political Campaign:
Opposition research team, ???
-Private Sector: CMO, ???
Beneficiaries
Mission AchievementMission Budget/Costs
Buy-In/Support
Deployment
Value
Proposition
Key Activities
Key Resources
Key Partners
Development Timeline Next Steps
20192017 2018 2020 20212016
Q1 Q2 Q1 Q2 Q1 Q2 Q1 Q2 Q1 Q2Q1 Q2 Q3 Q4 Q3 Q4 Q3 Q4 Q3 Q4 Q3 Q4 Q3 Q4
R&D/Dual-Use
Research
Join C5/Apply for OTA OTA Iterative Testing
Research Systems Integration
Burden
Hire Initial Data
Scientists/Engineers
Requirement Dev
User
Onboarding
Seed:
$1.2M
Series A:
$4.75M
Series B:
$20M
Submit
Proposals
Demonstrate Support Capability
System Threat Assessment
Hire Onboarding
Firm
Verify Requirement Compliance
Joint Staff Approval
R&D: Engineering + Design $332,000 Sales $160,000
R&D: Data science $300,000 Support $70,000
QA $80,000 Source Data $250k month (est)
Office, travel, admin, HR $250,000 Crowd labor $90k month (est)
Investment Readiness Level
IRL 1
IRL 4
IRL 3
IRL 2
IRL 7
IRL 6
IRL 5
IRL 8
IRL 9
First pass on MMC w/Problem Sponsor
Complete ecosystem analysis petal diagram
Validate mission achievement (Right side of canvas)
Problem validated through initial interviews
Prototype low-fidelity Minimum Viable Product
Value proposition/mission fit (Value Proposition Canvas)
Validate resource strategy (Left side of canvas)
Prototype high-fidelity Minimum Viable Product
Establish mission achievement metrics that matterTeam Assessment :
IRL 4

Narrative Mind Lessons Learned H4D Stanford 2016

  • 1.
    Original Problem: “We seekto develop tools that will optimize discovery and investigation of adversary communication trends on social media, allowing ARCYBER and others to more efficiently respond and mitigate threats posed by enemy messaging.” Team Narrative Mind Sponsor: US Army Cyber Command (ARCYBER) 100 Ten weeks of interviews later... Current Status: We learned a lot about this this space and the acquisition process.
  • 2.
    Show similar commercial tools--hearfeedback The Journey Week # EmotionalState 0 1 2 3 4 5 6 7 8 9 Suggestions for MVP switches each week Hashtag co- occurrence framework ARCYBER visit, spec of final MVP Unclear response from stakeholders.
  • 3.
    How do webegin? Our sponsor gave us a lot of freedom to explore.
  • 4.
    Initial Mission ModelCanvas (Week 1) - Gnip/Twitter - CrowdFlower, Samasource, or Mechanical Turk - Pre-existing social media service and micro- labor aggregators - Optimize workflow for social media analysts - Expedite categorization of social media content. - Use MechanicalTurk to crowdsource categorization of content. - - Algorithmic virality predictor ARCYBER wants to derive “meaning” Primary: Intelligence analysts receive a better platform. - Help intelligence analysts: receive cleaner, pre-categorized data, - Architecture that can support massive concurrent data aggregation and analysis. E.g. Storm/Hadoop. - Testing with analysts - MechanicalTurk or crowdsourcing labor (microtasks) - UI Development/Testing with CYBERCOM/ARCYBER analysts. - Software Development - Access to Twitter firehose (Stanford academic license) - Individual Analysts - ARCYBER - Continued partnership with crowdsourcing firms, CrowdFlower, Samasource, etc. Beneficiaries Mission AchievementMission Budget/Costs Buy-In/Support Deployment Value Proposition Key Activities Key Resources Key Partners - Categorize SM posts by content Week 1
  • 5.
    How do webegin? We learned about a wide variety of problems. Weeks 1-3
  • 6.
    DoD/Gvt. social mediapresence is weak. Account bans and multiple aliases across different networks make IDs hard to track Can’t monitor who views dark-web content. Fail to understand which narratives are the most salient. No baseline for monitoring/aggregating use of tech; Language/culture experts aren’t able to work at scale. Scale of social media makes manual efforts painful. Significant data management overhead. Can’t determine actual scale Problems Weeks 1-3
  • 7.
    How did werespond? We brainstormed a lot of MVP’s. Weeks 1-3
  • 8.
  • 9.
  • 10.
    Information Overload How dowe conceptualize and organize the challenges our users face? Week 4
  • 11.
    What kinds ofproblems are there? Global Tweet-level Awareness Response Mapping the Problem Space Week 4
  • 12.
    What kinds ofproblems are there? Global Tweet-level Awareness Response Mapping the Problem Space Week 4
  • 13.
    What kinds ofproblems are there? Global Tweet-level Awareness Response Mapping the Problem Space Week 4
  • 14.
    Mapping the ProblemSpace Week 4 What kinds of problems are there? Global Tweet-level Awareness Response
  • 15.
    Mapping the ProblemSpace Awareness Response Week 4
  • 16.
    Mapping the ProblemSpace Global Tweet-level Awareness Response Week 4
  • 17.
    Mapping the ProblemSpace Global Tweet-level Awareness Response IO Org Chart No baseline for monitoring/aggregating use of tech Automatic Narrative Detection Language/culture experts aren’t able to work at scale Important Event Predictor Preempt real world events Persistent ID-Alias tracker Account bans and multiple aliases across different networks make IDs hard to track Site Scraper Need more cached information access Expedited Content Categorization Scale of social media makes manual efforts painful Bot Detector Can’t determine actual scale of info. Virality Predictor Understand which narratives are the most salient
  • 18.
    Company F Persistent ID-Aliastracker Existing Products Global Tweet-level Awareness Response IO Org Chart No baseline for monitoring/aggregating use of tech Automatic Narrative Detection Language/culture experts aren’t able to work at scale Company C Virality Predictor Understand which narratives are the most salient Counter-Narrative Generator DoD/Gvt. social media presence is weak Company E Site Scraper Need more cached information access Company D Expedited Content Categorization Scale of social media makes manual efforts painful Company B Bot Detector . Company A Important Event Predictor Preempt real world events
  • 19.
    Opportunities Global Tweet-level Awareness Response Adversary IOOrg Chart No baseline for monitoring/aggregating adversary use of tech; cyber targeting Automated Narrative Detection Language/culture experts aren’t able to work at scale Open Opportunities
  • 20.
    Our Beneficiaries How canwe understand our users and the environment in which they operate? Week 5-6
  • 21.
    MMC - Week6 - Track how groups use technology over time. - Gnip/Twitter/Facebook - CrowdFlower, Samasource, or Mechanical Turk - Third-party access platforms for social media -Data visualization -Content analysis platforms Primary ARCYBER -Bg. General (decision maker) -MAJ/LTC/COL (operational plan) -Analysts/Operators (actionable insights) COCOMs -General (decision maker) -MAJ/LTC/COL -Analyst/Operator Secondary Political Campaigns -Campaign managers -Supporters Consumer Brands -CMO -Public Relations Team - Optimize workflow for social media analysts. -Deliver insights to commanders about online environment. -Insights into responses against narratives -Detect narratives emerging in real time -Early warning on emerging brand issues ??????????? -Enable faster problem awareness to problem response times for decision-makers across organizations- UI Development/Testing with ARCYBER analysts. - Software Development - Research aggregation - Access to Twitter firehose or API - Local language speaking crowdsourcing staff. - Accurate testing for intercoder reliability - ARCYBER: Bg. General, LTC, Strategic Initiatives Group, OTA, Purchasing PMs, End-User operator) -COCOMs: OTA, Operators, Purchasing PMs -Political Campaign: Opposition research team, ??? -Private Sector: CMO, ??? Beneficiaries Mission AchievementMission Budget/Costs Buy-In/Support Deployment Value Proposition Key Activities Key Resources Key Partners Week 6
  • 22.
    Value Proposition Canvas Products &Services Web/Desktop Application Act on reports and plans generated - Little understanding of ground-level nuances - Little understanding of proliferating platforms Customer Jobs Gains Pains Gain Creators Pain Relievers - Narrative detection and topic categorization - Global awareness - Reduce uncertainty of decision making - Add methodological rigor to ARCYBER’s operations Awareness ARCYBER - BG General (Decision Maker)
  • 23.
    Bg. General (decision maker) ❏Quickly understand key thematic points of organization's use of social media put out by intelligence briefs. MAJ/LTC/COL (operational planer) ❏ Determine what types themes are rising in popularity and better identify type of response Analysts/Operators (actionable insights) ❏ New movements can be understood and tracked with less direct cooperation of experts. Mission Achievement ARCYBER Big Picture Success Analogy: “Most COCOMs and IO shops spend their whole day looking for a needle in a haystack: a user, a post, an IP address. For narrative-level awareness, we need a strategy that helps us divide the haystack into a bunch of smaller haystacks that don’t all look like same damn pile of hay.”
  • 24.
    Low-Fidelity MVP Week 7 Howcan we incrementally improve ARCYBER’s existing workflow?
  • 25.
    Customer Quote “Things arechanging everyday. We need something that can help us with our long-term strategy, regardless of how many times we have to adjust our execution.” Week 7
  • 26.
  • 27.
    Research Dataset Dataset Features: ●600k Unique Tweets ● Spanning October 2015 to May 2016 ● 200k Unique Hashtag Combinations Procedure: 1. Merged records into frequency table of hashtag co- occurrences. 2. Manually coded 1300 most frequent hashtag sets 3. Visualized volume of these hashtag sets over time as related to big-picture “narrative”. Key Findings: ● Process is reasonably scalable. ● Could be implemented quickly by ARCYBER to supplement workflow. ● Deciding on narrative categories is difficult. ● Need to further condition input tweets to specific groups. ● How does hashtagged traffic compare to total traffic? Week 7
  • 28.
    Low-Fidelity MVP Weeks 6-8 Howwould any of these MVPs ever make it into the hands of our sponsors?
  • 29.
    First Steps: OTA Provinga prototype and informing a requirement. Innovation Challenge Initial Award Testing Evaluation ~$5M 1. Industry Day 2. Reqs. synopsis 3. Submit white papers 4. Evaluate papers 5. Proposals selected. 6. Technical discussion. Original requirements synopsis modified. 1. ACT office has personnel working with testers from all ranks. 2. Army Cyber Battle Lab involved for integration/concepts. Budget Adoption 1. Requirement written. 2. Companies apply. 3. If OTA company is selected, may skip as far as MS-B. 4. Follow FAR/JCIDS. 1. Problem requirements are changed iteratively. 2. Phase objectives and timeline are flexible. Weeks 6-8 Traditional and non-traditional contractors must partner.
  • 30.
    MMC - Week7 - Track how groups propagate narratives with co-occurring hashtags. - Gnip/Twitter/Facebook - CrowdFlower, Samasource, or Mechanical Turk - Pre-existing social media service and micro- labor aggregators - Third-party access platforms for social media -Data visualization -Content analysis platforms (Sens.ai, Leidos) Primary ARCYBER -Bg. General (decision maker) -MAJ/LTC/COL (operational plan) -Analysts/Operators (actionable insights) COCOMs -General (decision maker) -MAJ/LTC/COL -Analyst/Operator Secondary Political Campaigns -Campaign managers -Supporters Consumer Brands -CMO -Public Relations Team OTA in parallel for product development Create dual-use demand w/ PR and political campaigns - Optimize workflow for social media analysts. -Deliver insights to commanders about online environment. -Insights into responses against broadcasting narratives -Enable faster problem awareness to problem response times for decision-makers across organizations - MechanicalTurk or crowdsourcing labor (microtasks) - UI Development/Testing with ARCYBER analysts. - Software Development - Research aggregation? - Access to Twitter firehose or API - Local language speaking crowdsourcing staff. - Accurate testing for intercoder reliability - ARCYBER: Bg. General, LTC, Strategic Initiatives Group, OTA, Purchasing PMs, End-User operator) -COCOMs: OTA, Operators, Purchasing PMs -Political Campaign: Opposition research team, ??? -Private Sector: CMO, ??? Beneficiaries Mission AchievementMission Budget/Costs Buy-In/Support Deployment Value Proposition Key Activities Key Resources Key Partners
  • 31.
    Development Timeline NextSteps 20192017 2018 2020 20212016 Q1 Q2 Q1 Q2 Q1 Q2 Q1 Q2 Q1 Q2Q1 Q2 Q3 Q4 Q3 Q4 Q3 Q4 Q3 Q4 Q3 Q4 Q3 Q4 R&D/Dual-Use Research Join C5/Apply for OTA OTA Iterative Testing Research Systems Integration Burden Hire Initial Data Scientists/Engineers Requirement Dev User Onboarding Seed: $1.2M Series A: $4.75M Series B: $20M Submit Proposals Demonstrate Support Capability System Threat Assessment Hire Onboarding Firm Verify Requirement Compliance Joint Staff Approval R&D: Engineering + Design $332,000 Sales $160,000 R&D: Data science $300,000 Support $70,000 QA $80,000 Source Data $250k month (est) Office, travel, admin, HR $250,000 Crowd labor $90k month (est)
  • 32.
    Investment Readiness Level IRL1 IRL 4 IRL 3 IRL 2 IRL 7 IRL 6 IRL 5 IRL 8 IRL 9 First pass on MMC w/Problem Sponsor Complete ecosystem analysis petal diagram Validate mission achievement (Right side of canvas) Problem validated through initial interviews Prototype low-fidelity Minimum Viable Product Value proposition/mission fit (Value Proposition Canvas) Validate resource strategy (Left side of canvas) Prototype high-fidelity Minimum Viable Product Establish mission achievement metrics that matterTeam Assessment : IRL 4

Editor's Notes

  • #6  We continued learning…
  • #9 (3:05)
  • #15 (4:00)
  • #24 By mapping out these mission achievement goals for the various tiers of ARCYBER-- from the analyst who generates actionable insights, to the major making the plans, to the general who makes the final call--we were able to better conceptualize what a solution to this problem might look like
  • #25 Shifting our focus back to analyst level…
  • #26 7:35