32
TK Keanini Distinguished Engineer June 2018 And the Peak of Inflated Expectations Machine Learning

Machine Learning and the peak of Inflated Expectations · “Advanced Threats “100% predictive ... Common Techniques Unsupervised Learning ... If you can phrase a problem as 'we

Embed Size (px)

Citation preview

TK KeaniniDistinguished EngineerJune 2018

And the Peak of Inflated Expectations

Machine Learning

Time

Expecta

tions

Innovation Trigger Peak of Inflated Expectations Trough of Disillusionment Slope of Enlightenment Plateau of Productivity

Sourc

e: G

artn

er (J

uly

, 20

17

)

M A C H I N E L E A R N I N G

Gartner Hype Cycle for Emerging Technologies | 2017

Vendors Got Us Here

“Our machines detect threats others cannot”

“100% predictive”“Advanced Threats are no match for A.I.”

© 2018 Cisco and/or its affiliates. All rights reserved. Cisco Public

Silver Bullet Marketing

No Explanation or Discussion

Limited Guidance

How We Disservice Machine Learning

© 2018 Cisco and/or its affiliates. All rights reserved. Cisco Pulbic

MACHINE LEARNING

What it is

“Field of study that gives computers the ability to learn without being explicitly programmed.”Arthur Samuel’s definition of machine learning in 1959

bayesianbayesian

regressionregression

machine learningalgorithmsmachine learning

algorithms

regularizationregularizationclusteringclustering

neural networkneural networkdeep learningdeep learning

ensembleensemblerule systemrule system

decision treedecision tree

instance basedinstance based

dimensionality reductiondimensionality reductionN E R D A L E R T

Let’s define the helpful data science terms

classifierclassifierground truthground truth

bayesianbayesian

regressionregression

machine learningalgorithmsmachine learning

algorithms

regularizationregularizationclusteringclustering

neural networkneural networkdeep learningdeep learning

ensembleensemblerule systemrule system

decision treedecision tree

instance basedinstance based

dimensionality reductiondimensionality reduction

classifierclassifierground truthground truth

© 2018 Cisco and/or its affiliates. All rights reserved. Cisco Public

Machine Learning

Artificial Intelligence

Machine Learning

Supervised Learning

Unsupervised Learning

Reinforcement Learning

The Big Picture

© 2018 Cisco and/or its affiliates. All rights reserved. Cisco Public

Machine Learning

Common Techniques

Unsupervised Learning

Reinforcement Learning

Supervised Learning

When you know the question you are trying to ask and have examples of it being asked and answered correction

“The other” categoryTrial and error behavior effective in game scenarios

You don't have answers and may not

fully know the questions

© 2018 Cisco and/or its affiliates. All rights reserved. Cisco Public

E

Supervised Learning Unsupervised Learning Other(Reinforcement Learning, etc.)

75% 15% 10%

© 2018 Cisco and/or its affiliates. All rights reserved. Cisco Public

a

Use in combination with Machine Learning

What did we do before Machine Learning?

Simple Pattern Matching

Statistical Methods Rules and First Order Logic (FoL)

© 2018 Cisco and/or its affiliates. All rights reserved. Cisco Public

MACHINE LEARNING

Techniques

“Field of study that gives computers the ability to learn without being explicitly programmed.”

“Field of study that gives computers the ability to be implicitly programmed.”

Translation

Classifier Prediction

© 2018 Cisco and/or its affiliates. All rights reserved. Cisco Public

Training Data

Training Classifiers

Machine Learning Algorithm

New Data

Ground Truth Used in Supervised LearningThe 'Ground Truth' is the pairing of example questions and answers.

If you can phrase a problem as 'we know this is right, learn a way to answer more questions of this type'.

Success depends greatly on the dataset expressing the Question -> Answer mapping.

© 2018 Cisco and/or its affiliates. All rights reserved. Cisco Public

MACHINE LEARNING

Pitfalls

One Size Does Not Fit All

Other ML Application Security

N E R D A L E R T

Warning: Success in one domain does not guarantee success in another© 2018 Cisco and/or its affiliates. All rights reserved. Cisco Public

What Is At Stake Matters

Because you watched Deadpool, you might like…

Deadpool X-Men: First Class The Flash Captain America: The First Avenger

© 2018 Cisco and/or its affiliates. All rights reserved. Cisco Public

How did you come to that conclusion?

“The Explainability Problem”

Normal WorkflowCFO daily calendar

Irregular ActivityML detects “suspicious”

activity and suggests remediation

QuarantinedHowever, ML cannot

articulate *why* it wants to remediate

Loss of time and resources

© 2018 Cisco and/or its affiliates. All rights reserved. Cisco Public

MACHINE LEARNING

For Security

PrecisionWhen my classifier predicts an instance in a certain class, how often does the instance belong to that class?

AccuracyHow often does my classifier give me the correct answer?

How We Know Machine Learning is Working

N E R D A L E R T

Root mean square error & Logical RegressionTranslation: On average, how far away are my predictions from what we later know to be true values?

© 2018 Cisco and/or its affiliates. All rights reserved. Cisco Public

Why is Machine Learning so useful in Security?

Evolving SecurityThe security domain is always evolving,

has a large amount of variability,

and is not well-understood

StaticWith limited variability or is

well-understood

© 2018 Cisco and/or its affiliates. All rights reserved. Cisco Public

Insider Threats and Behavioral Security Analytics

DetectingThrough novelty and outliers

AttackersThey’re not breaking in,

they are logging in

EventsTurn weak signals into a

strong one

© 2018 Cisco and/or its affiliates. All rights reserved. Cisco Public

Classify the Observable World and Infer the Rest

Normal ActivityWeird Stuff(but not threat related)

Threat Actor Activity

© 2018 Cisco and/or its affiliates. All rights reserved. Cisco Public

Billions of connections

• Statistical Methods

• Information-Theoretical Methods

• 70+ Unsupervised Anomaly Detectors

• Dynamic Adaptive Ensemble Creation

• Multiple-Instance Learning

• Neural Networks

• Rule Mining

• Random Forests

• Boosting

• ML: Supervised Learning

• Probabilistic Threat Propagation

• Graph-Statistical Methods

• Random Graphs

• Graph Methods

• Supervised Classifier Training

Anomaly Detectionand Trust Modeling

Event Classificationand Entity Modeling

RelationshipModeling

Cascade of specialized layers of Machine Learning algorithms

Multi-layer Analytical Pipeline

© 2018 Cisco and/or its affiliates. All rights reserved. Cisco Public

5

Security that Shows its Work

5

Oct.

3

3 Spam tracking#CSPM02

New

3

Oct.

4

C&

C u

rl

8 Information Stealer#CDCH01

Oct.

15

Anom

alo

us

htt

p

37

Oct.

16

Heavy

uplo

ader

Dro

pbox.c

om

78

Oct.

25

Oct.

28

Malic

ious

htt

p

Recurr

ing

8Malware: salityDec. 9 | 28 days

© 2018 Cisco and/or its affiliates. All rights reserved. Cisco Public

Measure the Right Things

Efficacy of the Assertions

True/False Positive

True/False Negative

Root Mean Squared Error

Overfitting/Undefitting

© 2018 Cisco and/or its affiliates. All rights reserved. Cisco Public

Measure the Right Things

© 2018 Cisco and/or its affiliates. All rights reserved. Cisco Public

Conclusion

How are you applying Machine Learning in your product and why?

How do you measure its effectiveness?

What to Ask Your Vendor

Regarding supervised learning, what are you using for ’ground truth’?

What non-machine learning are you using and why?

What papers or open-source have you published regarding your analytics?

For the ML based assertions, what entailments are provided?

© 2018 Cisco and/or its affiliates. All rights reserved. Cisco Pulbic

Be Pragmatic

Entailments

Success is Domain Specific

Analytical pipeline, over single technique

A Good Machine Learning Approach

Measure helpfulness, not mathematical accuracy

© 2018 Cisco and/or its affiliates. All rights reserved. Cisco Public

N E R D A L E R T

Thank you!