Stealing Machine Learning Models via Prediction APIs

Stealing Machine Learning Models via Prediction APIs
1
Florian Tramèr , Fan
2
Zhang , Ari
3
Juels , Michael
4
Reiter , Thomas
3
Ristenpart
1EPFL, 2Cornell, 3Cornell Tech, 4UNC
Goals
§ Machine learning models may be
deemed confidential due to
§ Sensitive training data
§ Commercial value
§ Use in security applications
§ In practice, ML models are deployed
with public prediction APIs.
§ We show simple, efficient attacks
that can steal the model through
legitimate prediction queries.
Approach cont.
Decision Tree: Path-Finding Attacks
§ We propose a new Path-Finding attack
§ Exploited the ability to query APIs with
incomplete inputs.
§ Also apply to regression trees.
Results
Approach
Model Extraction against MLaaS
ML#service#
Data#owner#
f (x1 )
…#
Train#
model##
DB#
x1
Extrac3on#
adversary#
xq
fˆ
ü Tables shows the number of prediction
queries made to the ML API in an attack
that extracts a 100% equivalent model:
f (xq )
LR and MLP: Equation-Solving
Service Model Data set
Amazon LR
Digits
LR
Adult
BigML
DT
German Credits
DT
Steak Survey
Makinguseoftheconfidencevalues.
§ Logistic Regression: π’˜ β‹… 𝒙 = 𝜎 𝑓 𝒙
§ Multiclass LR (MLR) and Multilayer
Perceptron (MLP):
𝜎(𝑖, π’˜πŸ β‹… 𝒙, … , π’˜π’„ β‹… 𝒙) = 𝑓. (𝒙)
Model
𝜎(𝑖, 𝜢0 β‹… πœ… 𝒙, 𝝉 , … , 𝜢4 β‹… πœ… 𝒙, 𝝉 )
= 𝑓. (𝒙)
Makinguseofonlytheclasslabel.
§ Retraining with uniform queries
§ Line-search retraining
§ Adaptive retraining
http://silver.web.unc.edu
Time (s)
70
149
632
2,088
Success of equation-solving attacks
§ Kernelized LR:
SVM: Retraining
Queries
650
1,485
1,150
4,013
Unknowns
Softmax
530
OvR
530
MLP
2,225
Queries
1-R_test
1-R_unif
Time (s)
265
530
265
530
2,225
4,450
99.96%
100.00%
99.98%
100.00%
98.68%
99.89%
99.75%
100.00%
99.98%
100.00%
97.23%
99.82%
2.6
3.1
2.8
3.5
168
196
Training data extraction
Training data:
Recovered:
Cloud Security Horizons Summit, March 2016