SESSION ID:

#RSAC

Multiparty Computation and Application

CRYP-F01

Eleftheria MakriLecturer/Researcher Imec-COSIC, KU Leuven, BE & Saxion University of Applied Sciences, NL @MakriEleftheria

EPIC: Efficient Private Image Classification (or: Learning from the Masters) E. Makri, D. Rotaru, N. P. Smart, F. Vercauteren

#RSAC

EPIC: Efficient Private Image Classification

4

#RSAC

5

Feature Extraction

#RSACTransfer Learning Feature Extraction (or: Learning from the Masters)

6

#RSAC

7

#RSAC

EPIC Security

8

Active Security vs. Passive Security

EPICAll other protocols in the related work

#RSAC

9

Step 1: Create the ML model

Dragos Rotaru 9

Inception–v3 CNN

Linear SVM

Alice

#RSAC

10

Step 2: Alice secret shares the ML model

𝑆𝑆𝑆𝑆𝑆𝑆𝐴𝐴

𝑆𝑆𝑆𝑆𝑆𝑆𝐶𝐶

AliceBob

Charlie

#RSAC

11

Step 2: Alice secret shares the ML model

𝑆𝑆𝑆𝑆𝑆𝑆𝐴𝐴

𝑆𝑆𝑆𝑆𝑆𝑆𝐶𝐶

𝐾𝐾1 𝐾𝐾2 𝐾𝐾3+ + =Dragos Rotaru

𝑆𝑆𝑆𝑆𝑆𝑆𝐴𝐴 𝑆𝑆𝑆𝑆𝑆𝑆𝐵𝐵 𝑆𝑆𝑆𝑆𝑆𝑆𝐶𝐶

AliceBob

Charlie

#RSAC

12

Step 3: Bob extracts features

Dragos Rotaru 12

Inception–v3 CNN

Features 𝑋𝑋

Bob

#RSAC

13

𝑆𝑆𝑆𝑆𝑆𝑆𝐴𝐴

𝑆𝑆𝑆𝑆𝑆𝑆𝐶𝐶

Dragos Rotaru

Step 4: Bob secret shares features

AliceBob

Charlie

#RSAC

14

𝑆𝑆𝑆𝑆𝑆𝑆𝐴𝐴

𝑆𝑆𝑆𝑆𝑆𝑆𝐶𝐶

Dragos Rotaru

Step 4: Bob secret shares features

𝑋𝑋𝐴𝐴

𝑋𝑋𝐶𝐶

𝑋𝑋𝐵𝐵

𝐾𝐾1 𝐾𝐾2 𝐾𝐾3+ + =Dragos Rotaru

𝑋𝑋𝐴𝐴 𝑋𝑋𝐵𝐵 𝑋𝑋𝐶𝐶 CNN-Feat( )

AliceBob

Charlie

#RSAC

15

𝑆𝑆𝑆𝑆𝑆𝑆𝐴𝐴 𝑆𝑆𝑆𝑆𝑆𝑆𝐶𝐶

Step 5: Parties use MPC to help Charlie compute label of SVM-Alice(Bob-Image)

𝑋𝑋𝐴𝐴 𝑋𝑋𝐶𝐶

𝑋𝑋𝐵𝐵

Alice

Bob

Charlie

#RSAC

16

𝑆𝑆𝑆𝑆𝑆𝑆𝐴𝐴 𝑆𝑆𝑆𝑆𝑆𝑆𝐶𝐶

Step 5: Parties use MPC to help Charlie compute label of SVM-Alice(Bob-Image)

𝑋𝑋𝐴𝐴 𝑋𝑋𝐶𝐶

𝑋𝑋𝐵𝐵

***“Florist”***Alice

Bob

Charlie

#RSAC

Video of the Demo of our work will appear here

17

#RSAC

EPIC Performance – Simple Variant

18

Computation Cost Communication Cost

0

50

100

150

200

250

300

CIFAR-10(88.8% accuracy)

MIT-67(72.2% accuracy)

Caltech-101(91.4% accuracy)

Communication (MB)

Offline Online Total

0

0.5

1

1.5

2

2.5

3

3.5

4

CIFAR-10(88.8% accuracy)

MIT-67(72.2% accuracy)

Caltech-101(91.4% accuracy)

Runtimes (s)

Offline Online Total

#RSAC

Performance of the state-of-the-art private image classification

19

Computation Cost Communication Cost

0.001

0.01

0.1

1

10

100

1000

MiniONN*(81.61% accuracy)

Gazelle**(81.61% accuracy)

EPIC(88.8% accuracy)

Runtimes (s)

Offline Online Total

0.1

1

10

100

1000

10000

MiniONN*(81.61% accuracy)

Gazelle**(81.61% accuracy)

EPIC(88.8% accuracy)

Communication (MB)

Offline Online Total

* Jian Liu, Mika Juuti, Yao Lu, N. Asokan. Oblivious Neural Network Predictions via MiniONN Transformations. In Proceedings of the 2017 ACM SIGSAC Conference on Computer and Communications Security (pp. 619-631). ACM.** Chiraag Juvekar, Vinod Vaikuntanathan, and Anantha Chandrakasan. GAZELLE: A low latency framework for secure neural network inference. In 27th USENIX Security Symposium (USENIX Security '18), Baltimore, MD, 2018. USENIX Association.

#RSAC

EPIC Efficiency Gain over the state-of-the-art

20

EPIC vs. Gazelle1 on CIFAR-10: – 34 times faster runtime;– 50 times improvement of communication cost; – 7% higher classification accuracy.

EPIC vs. Gazelle1 with the same accuracy: – 700 times faster runtime; – 500 times improvement of communication cost.

1 Chiraag Juvekar, Vinod Vaikuntanathan, and Anantha Chandrakasan. GAZELLE: A low latency framework for secure neural network inference. In 27th USENIX Security Symposium (USENIX Security '18), Baltimore, MD, 2018. USENIX Association.

#RSAC

Now what?

21

What would transform EPIC to a LEGENDARY solution? – Maintain security – Maintain or increase efficiency – Increase accuracy!

Any ideas on how to do this (using MPC)? – Talk to me during the break, or– Contact me offline at: eleftheria.makri@esat.kuleuven.be

mailto:eleftheria.makri@esat.kuleuven.be

Multiparty Computation and Application EPIC: Efficient Private Image Classification (or: Learning from the Masters) Slide Number 3EPIC: Efficient Private Image Classification Slide Number 5Transfer Learning Feature Extraction (or: Learning from the Masters) Slide Number 7EPIC SecurityStep 1: Create the ML modelStep 2: Alice secret shares the ML modelStep 2: Alice secret shares the ML modelStep 3: Bob extracts featuresSlide Number 13Slide Number 14Slide Number 15Slide Number 16Video of the Demo of our work will appear here EPIC Performance – Simple Variant Performance of the state-of-the-art private image classificationEPIC Efficiency Gain over the state-of-the-art Now what? Slide Number 22