Presentation given at the "Interchain Conversations" conference in Berlin (Full Node, June 2019)

1. Nicolas Pinto
Interchain Conversations, Berlin | June 2019
edition
Buzzwords
(a few more buzzwords here)
buzzword

2. Nicolas Pinto
Interchain Conversations, Berlin | June 2019
edition
”AI” for Cosmos Security
Icarus: Deep (Learning) Explorations of the Cosmos SDK*
*warning: very preliminary work

3. Goal
• ”AI”-driven Security Audit for Blockchain Projects
• Swim in buzzwords and equations*
• Spark interest around a non-obvious topic
• Present encouraging preliminary results
• Bring two different cultures together
• Get you involved!

4. Outcome

5. Outline
1. Intro: AI4BC / BC4AI
2. Methods: KISS
3. Results: ZBRA
4. Discussion: YOU

6. 1. Intro: ai4bc / bc4ai
Buzzword Buzzword, Inc.
(longer than usual)

7. Motivation...

8. What?
• Applying “AI” / (Deep) Machine Learning to hunt
for vulnerabilities in Blockchain projects.
• Think of DeepBlue, Watson, AlphaGo, AlphaStar
• But for Bugs!
…

9. — Who?
• Thomas France (CEO)
• Nicolas Pinto (CTO)
• Both interested in “AI/ML” and “BC”

10. — Who?
• Thomas France (CEO):
• Co-founder @ Bitcoin House in Paris
• Co-founder @ Ledger
• Angel investor & early contributor in crypto
• Left Ledger in January 2019 to explore new
projects around scalability, security and
privacy

11. — Who?
• Nicolas Pinto (CTO):
• 2006-2010: Neuroscience/AI PhD @ MIT
• 2009: Enthusiast @ Bitcoin
• 2010-2012: Neuroscience/AI @ MIT/Harvard
• 2012-2014: Co-founder @ Perceptio (Mobile Deep Learning)
• 2014-2018: Ninja, Research Director @ Apple
• 2018-now: Blockchain for AI <> AI for Blockchain

12. 13 years in 5 minutes

13. Back in 2006

14. Neuroscience
DiCarlo Lab @ MIT
Jim DiCarlo
David Cox

15. The Approach
Reverse and Forward Engineering Intelligence
(Neuro)

16. The Approach
Reverse and Forward Engineering the Brain
Build
Artificial System
FORWARDREVERSE
Study
Natural System

17. Reverse Engineering
TheVisual Cortex
brain = 20 petaflops ?!

18. PeakGFLOP/s
Computationalpower
GPUs
CPUs
DIY GPU pr0n (since 2006) Sony Playstation 3s (since 2007)
The blessing of GPUs 2006

19. Build your own!
2006

20. Build your own!
Sony Playstation 3s (since 2007)
2006

21. MIT (2008-2009)
Harvard (2010-2012)
Teaching our craft
Sharing as early as possible, empowering others…

22. @end: 100s of GPUs (for free!)
2010-2012 @ MIT/Harvard

23. Pinto,Doukhan,DiCarlo,CoxPLoS2009
n
e
nd-
en
tes
ates
t
es
states,
ap of
is
neigh-
low, the
becomes
o the
nd Rabi
onal states
ay lead to
kely a
nformation
009).
tein expression p
modium that were isolated from the
patients; they identified about 100 proteins,
some of which had not been found in laboratory
cultures and could make promising drug or vac-
cine targets. — HP
Proteomics Clin. Appl. 3, 1314 (2009).
N E U R O S C I E N C E
The Next Top Model
Consumers may be familiar with high-end
graphic processing components in video game
consoles, such as the PlayStation3, or as a con-
sequence of outfitting personal computers
ordered online with NVIDIA graphics cards;
these advances
in hardware
have also
attracted the
attention of
procurement
officials in the
military serv-
ices. In the aca-
demic realm,
Pinto et al. have
harnessed the power
of clustered graphics
processors to assess the relative performance of
sion models of object recognition.
arallel processing
onJanuary4,2010
www.sciencemag.org
Downloadedfrom
C S
owave Manipulation
l lattices populated by neutral atoms are
candidate for storing quantum infor-
. Normally, internal degrees of freedom
s the hyperfine state are used to create
sic information unit, the qubit. However,
also possess motional degrees of free-
or example, the confinement of atoms in
EDITORS’CHOICE
Plasmodium in comparison to laboratory cul-
tures revealed differences in gene expression
profiles. Acharya et al. have analyzed the pro-
tein expression profiles of two species of Plas-
modium that were isolated from the blood of
patients; they identified about 100 proteins,
some of which had not been found in laboratory
cultures and could make promising drug or vac-
cine targets. — HP
Proteomics Clin. Appl. 3, 1314 (2009).
EDITED BY GILBERT CHIN AND JAKE YESTON
full control of quantum transport, likely a
necessity for processing quantum information
in this system. — JS
Phys. Rev. Lett. 103, 233001 (2009).
C E L L B I O LO G Y
In the Wild
Malaria is one of the most prevalent infectious
diseases and kills around 900,000 people per
year. It is caused by parasites of the genus Plas-
modium, which are transmitted to humans by
mosquitoes and enter red blood cells, causing
fever and, if left untreated, death. Human
pathogens of all kinds can develop resistance to
the most effective drugs, such as artemisinin, so
there is a constant need to identify new com-
pounds. Animal models of malaria have proven
problematic to establish, and most studies have
used laboratory cultures of human blood cells to
grow the parasites. While important insights
into the life cycle and pathogenic action of Plas-
modium have come from these in vitro studies,
a recent study of clinically isolated samples of
www.sciencemag.org SCIENCE VOL 327 1 JANUARY 2010
Pinto et al. h
harnessed th
of clustered
processors to
machine visi
The availabi
power at rea
explore, in 1
regions of p
ber of filters
They genera
were trained
during an un
then screene
versus plane
of orientatio
The top-rank
broadly acro
toughest rec
human faces
sophisticate
set of param
with high ob
vessel called the ductus arteriosus (DA)
allows blood to bypass the nonfunctional
fetal lungs by connecting the pulmonary
artery, which supplies blood to the lungs,
with the aorta, which supplies blood to the
rest of the body. This vessel normally closes
a day or two after birth, but in some new-
borns, it remains open and can lead to life-
threatening complications. Studying new-
born mice, Echtler et al. make the surprising
observation that platelets—cells noted for
their role in blood clotting—were recruited
to the lumen of the DA within 20 minutes
after birth of the mice; when platelet pro-
duction or function was disrupted, the DA
failed to close completely, leading to abnor-
mal patterns of blood flow. The recruited
platelets play a dual role in DA closure—by
forming a physical plug that seals the lumen
of the constricted DA and by altering the
behavior of other cell types involved in
blood vessel remodeling. — PAK
Nat. Med. 10.1038/nm.2060 (2009).

24. Did Google get inspired?

25. PERCEPTIO
The world’s most advanced
mobile-first deep learning startup*
*also still the only one, as far as we know! ;-)
2012

26. Mobile first !
establishing trust, betting against the cloud…

27. A.I. you can trust
protecting your privacy

28. More data, less power
when running intelligence right next to the sensor

29. 2014

31. Is winter back?
AlexNetNP’s PhD
(Dec 2010)
Perceptio
(Aug 2012)
Perceptio 2.0 ?
(Mar 2018)
Perceptio @ Apple
(Nov 2014)

32. PERCEPTIO 2.0
Decentralized AI (bc4ai)
2018

33. Peter Thiel
What do you disagree with
that almost everyone else
believes? ***
”
“

34. Peter Thiel
AI centralizes,
Blockchain decentralizes…”“

35. Decentralized AI (bc4ai)
• Move away from centralized & supervised AI
• Are we in the Matrix?
• Non-dystopian future:
decentralized & unsupervised (autonomous) AI
• Collective intelligence:
• Scalable
• Private
• Secure
(making smart contracts actually smart, but not only)

36. Too Early?
• Decentralization (blockchain) technology
may not be mature enough
• Being too early is like being wrong…
• You don’t want to be doing Neural Networks
in the 80s…

38. Can we help…
• … accelerate progress?
• AI for Blockchain ;-)
• Before bc4ai, let’s do ai4bc !

39. ai4bc
(AI for Blockchain*)

40. Help with scalability?
(AI for Blockchain)

41. NIPS 2011
Scalability?we did something similar in the past…
ai4gpu

42. Google (2018)
Scalability?Google does it now…

43. • Parameters without closed-form solutions
(e.g. block size :-)
• Transaction throughput (i.e. tx / s)
• Fee prediction (e.g. Bitcoin’s estimatesmartfee 2.0 ?)
• Block construction (e.g. UTXO set selection)
• Crypto implementation (e.g. ZKP)
• etc.
Optimize what?(for scalability)

44. Feedback(from many blockchain devs)
Security First !

45. Security ?!
• Scalability? not yet.. Security first!
• Security is paramount for Blockchain
• Any failure: huge reputational risk, if not killer

46. Why Cosmos?
(Mister Anderson)

47. Why Cosmos ?! Because…
• Cosmos community showed us the way…
• Thanks in particular to:
• Adrian Brink (TokenSummit’17),
• Chris Goes (DevCon’18),
• Jae Kwon (BUIDL’18)
• Ethan Buchman (BPASE’19)

48. has nice swag!
Cosmos

49. Why Cosmos ?! Because…
• PBFT & IBC are sexy:
• Blockchain of blockchains,
• Internet of blockchains,
• OS of blockchains?
• Already used by many exciting projects:
IRIS, LOOM, Binance DEX, Terra, etc.
• Platform of choice for bc4ai ?!

50. Of course: Cosmos!!
• Just launched!
• We are big fans of the project:
• Supporters/investors in ICO, All In Bits,
CastleNode, Cryptium Labs, etc.
• Game-of-Stakes “winner”: BouBouNode !

51. BouBouNode

52. Cosmos x Security
• Talking about Game-of-Stakes…
• Cosmos has taken security seriously since day one
• Jae talked about a custom Ledger for Tendermint
validators in 2015!
• Implementation of best-in-class practices
very early on
• Security-driven development
(slow like in hardware)

53. COSMOS CODE IS PUBLIC !

54. Cosmos x Security
• Afraid? Many audits! Bounty program!
• Active community of security researchers with a
wide range of expertise, for example:
• Certus.One: infrastructure
• B-Harvest: economics

55. Cosmos x Security
• Interchain Foundation very open to experimenting
• Already pushing the boundaries of testing in
blockchain…
• Simulator available!

56. 2. Methods: KISS
Icarus

58. Hype vs. Reality
“AI”

59. “AI” vs. IA
• “AI” (today):
• Skeptic: “AI = glorified guided random search
(almost brute force)”
• Centralized & disrupting/displacing jobs
• IA (tomorrow):
• IA = Intelligence Augmentation
• Decentralized & empowering devs :)

60. Reality: Blue vs. White vs.
Pajamas Collar

61. Icarus: Preliminary Ideas
• Practical:
• Deep Learning for Fuzzing Cosmos (buzzwords ;-)
• Long shots:
• Formal verification relaxation:
bridge from dynamic to static verification ?
• Source code analysis:
“hotspot” prediction / classification

62. Bug “hotspot” prediction
• Simulate dev/auditor
• Generalizable?

63. Interviews: the cast
• Tendermint family: Chris, Jae, Ethan, Zaki, Anton, Jessy, Alex
• Validators:
• Cryptium Labs: Adrian*
• CastleNode: Gautier
• Certus.One: Hendrik
• B-Harvest: Hyung
• Independent security consultants
• Future collaborators
(in blockchain, security, and AI spaces)
etc.

64. Interviews: pain points
• Simulator (random tests already in place)
• Maths / Types (how numbers are handled)
• Amino (serialization)
• Jepsen (distributed system tests)
• P2P / RPC
• Fairness (e.g. block proposer)
• Mempool (e.g. non-obvious mempool DDoS)
• Trust boundaries
• Crypto economics / financial distribution logics
etc.
little overlap!

65. Interviews: quotes
• “It’s easier to find bugs from the data than from
looking at the code”
(Hyung @ B-Harvest)
• “The simulator is only vaguely fuzzing at the
moment, it’s more like a sophisticated integration
test suite”
• “Transactions are too clean in the simulator”
(Hendrik @ Certus.One)

66. Interviews: “bug correlates”
• (Deep) Learning needs gradients
• Bugs are non differentiable ;-)
• We need proxies/surrogates = "bug correlates”
• Signals that tend to correlate with bugs,
alone or in high-order / complex interactions
• Which ones would you suggest?
F(X) = Y ~ Z
0
1

67. Interviews: “bug correlates”
• Resources usage:
• Memory (e.g. leaks)
• Go routines
• File descriptors
• Stack depth
• Gas consumed (surrogate for complex operations)
• Block time (e.g. w.r.t. blockchain size)
• Long running chains
• Complex cross-module interactions (i.e. hooks)
etc.

68. Getting our feet wet
We focused on:
• Amino (fuzzer already in place)
• Jepsen tests
• Math / types
• Simulator (integration tests) *

69. Simulator: flow
sim params
output stats848275874
random seed
fast slow

70. sim params275874
dom seed
fast

71. output stats
slow

72. Simulator: “AI”
• Goal: focus compute power on more “interesting”
runs (defined as “likely to produce a bug” ;)
• Use machine learning for “interesting”
classification
• Like spam !

73. Simulator: “AI”
sim params
output sta75874
m seed
fast slowfast
yes
no
AI magic*

74. Linus Pauling
(double Nobel Prize Winner)
If you want to have good ideas
you must have many ideas.”
“
Most of them will be wrong,
and what you have to learn is
which ones to throw away.
“
”

77. 3. Results: preliminary
(super fresh)

79. Brute Force Data Collection
• KISS, DIY, duck-taped…
• ML: training/validation/testing all intertwined
• Not necessarily statistically correct,
but it doesn’t matter here
• All simulation runs are valid
• Any jamming during data collection is valuable

80. SuperComputer
• At first: my laptop + 2x 2€/month nodes ;-)
• Then: 6x CX51 nodes on Hetzner (6x 30€/month)
• 6x 8vCPUs / 32GB
• 48 concurrent processes

81. #blocks
count Insights
most “die” early

82. #blocks
time(s) Insights

83. TADAAAA !
• Dead lock in invariant checks
• 0.5% of the FAIL runs are “time out” (24h !!!!)
• i.e. you need 200 runs to get one!
• after a while all CPUs get stuck
BUG FOUND

84. BUGS FOUND

85. Simulator Bugs
• Low hanging fruits first (predictable)
• BUG: simulator non-determinism (RNG ops)
BUG FOUND

86. Simulator Bugs
• BUG: faulty parameter distribution boundaries
BUG FOUND

87. Simulator Bugs
• BUG: faulty parameter distribution boundaries
BUG FOUND

88. Simulator Bugs
• BUG: faulty parameter distributions
BUG FOUND
BUG: 1/8000 chance of sum == 0 FIX: hack

89. SDK Bugs
• BUG: number handling
BUG FOUND
small?

90. SDK Bugs
• BUG: number handling
BUG FOUND
small?

91. SDK Bugs
• BUG: number handling
BUG FOUND

92. Coverage
• Coverage is trending down as devs add more
features…
MORE !
https://codecov.io/gh/cosmos/cosmos-sdk

93. Coverage
• Brute force “AI” is already helping…
MORE !

94. Coverage
• Coverage optim (+test_cover.sh): from 56.5% to 61.6%
• However, code coverage is easier and less valuable
than state coverage…
MORE !

95. Copper Cluster
• Collaboration w/ Prof. Graham Taylor @ University of
Guelph (AI friend but also Ethan’s M.S. advisor!)
• Copper: a real SuperComputer
• Compute: 577 CPUs / concurrent processes
(12X more than before)
• 24/7, just for us!?
• 10Ks+ simulations per day
• Maybe: academic paper with Graham’s research group?
MORE !

96. Future?
MORE !

97. 4. Discussion

98. Kaggle-like Competition?
• Dataset sharing: 76,655 simulations with
associated profiling data
• Idea: Kaggle-style bounty-driven competition?

99. Feature Requests
• Direct control of simulation parameters
• Like hyper-parameter optimization in deep learning
sim params
output stats
fast slow
AI magic*

100. Feature Requests
• Direct control of simulation parameters
DONE

101. Feature Requests

102. Feature Requests
• Bug “Bubbles”:
• ways to parametrize sim around critical bugs?
• and use these as seeds to explore around…

103. Future Work
• Full transaction control during the simulation
• “AI” Bugs to “AI” Fixes?
• Semi-automated bug reports
• Problem: reports are (human) time consuming
to “manicure”
• Solution: attention model for highlights / insights

104. Getting you involved!
• A lot of surface area to cover: F(X) = Y ~ Z
• Engage you as well as the AI community
• Come make Cosmos the most tested/robust
project in blockchain software ever !

105. Thanks
• Thomas France
• ICF, Tendermint: Ethan, Jae, Chris, Alex, Rigel, Zaki,
Anton, Jessy
• Interviewees: Adrian, Gautier, Hendrik, Hyung, Anil
• Current collaborators: Graham Taylor, Fernand Pajot
• Future collaborators: Joseph Turian*, Nicolas Poilvert,
Edouard Oyallon
• BouBou(Node)

106. Outcome

107. COME