Machine Learning in the presence of adversaries · Machine Learning in the presence of adversaries...
Transcript of Machine Learning in the presence of adversaries · Machine Learning in the presence of adversaries...
-
Machine Learningin the presence of adversaries
(joint work with Mika Juuti, Jian Liu, Andrew Paverd and Samuel Marchal)
N. Asokan http://asokan.org/asokan/@nasokan
-
2
Machine Learning is ubiquitous
The ML market size is expected to grow by 44% annually over next five yearsIn 2016, companies invested up to $9 Billion in AI-based startups
Machine Learning and Deep Learningis getting more and attention...
2[1] http://www.marketsandmarkets.com/PressReleases/machine-learning.asp[2] McKinsey Global Institute, ”Artificial Intelligence: The Next Digital Frontier?”
http://www.marketsandmarkets.com/PressReleases/machine-learning.asp
-
Machine Learning for security/privacy
3
Access Control Deception DetectionMarchal et al., “Off-the-Hook: An Efficient and Usable Client-Side Phishing Prevention Application” 2017. IEEE Trans. Comput.https://ssg.aalto.fi/projects/phishing/
https://ssg.aalto.fi/projects/phishing/
-
4
How do we evaluate ML-based systems?
Effectiveness of inference• measures of accuracy
Performance• inference speed and memory consumption
...
Meeting these criteria even in the presence of adversaries?
-
Security & privacy ofmachine learning
-
6
Which class is this?School bus
Which class is this?Ostrich
Szegedy et al., “Intriguing Properties of Neural Networks” 2014. https://arxiv.org/abs/1312.6199v4
Skip to robust adversarial examples
+ 0.1⋅ =
Adversarial examples
https://arxiv.org/abs/1312.6199v4
-
7
Which class is this?Panda
Goodfellow et al., “Explaining and Harnessing Adversarial Examples” ICLR 2015
Which class is this?Gibbon
+ 0.007⋅ =
Adversarial examples
-
8
Adversarial examples
8Papernot et al, “Practical Black-Box Attacks against Machine Learning”, ASIACCS 2017 https://arxiv.org/abs/1602.02697
https://arxiv.org/abs/1602.02697
-
9
Which class is this?Cat
Which class is this?Desktop computer
Athalye et al. “Synthesizing Robust Adversarial Examples”. https://blog.openai.com/robust-adversarial-inputs/
https://blog.openai.com/robust-adversarial-inputs/
-
1010Zhang et al, “DolphinAttack: Inaudible Voice Commands”, ACM CCS 2017 https://arxiv.org/abs/1708.09537
https://arxiv.org/abs/1708.09537
-
11Fredrikson et al. “Model Inversion Attacks that Exploit Confidence Information and Basic Countermeasures”, ACM CCS 2015. https://www.cs.cmu.edu/~mfredrik/papers/fjr2015ccs.pdf
Model inversion
https://www.cs.cmu.edu/%7Emfredrik/papers/fjr2015ccs.pdf
-
12
Model inversion
12Hitaj et al. “Deep models under the GAN: information leakage from collaborative deep learning”, ACM CCS 2017. https://arxiv.org/pdf/1702.07464.pdf
https://arxiv.org/pdf/1702.07464.pdf
-
13
Machine Learning pipeline
Data owners
Analyst
𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇
𝐿𝐿𝑇𝑇𝐿𝐿𝐿𝐿
ML model Client
Prediction Service Provider
API𝐷𝐷𝑇𝑇𝑎𝑎𝑇𝑇𝐿𝐿𝑇𝑇𝑎𝑎
Where is the adversary? What is its target?
-
14
Speed limit 80km/h
𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇
𝐿𝐿𝑇𝑇𝐿𝐿𝐿𝐿
Compromised input
Data owners
Analyst
ML model
Prediction Service Provider
API
Dang et al., “Evading Classifiers by Morphing in the Dark”, ACM CCS ’17. https://arxiv.org/abs/1705.07535Evtimov et al., “Robust Physical-World Attacks on Deep Learning Models”. https://arxiv.org/abs/1707.08945Zhang et al., “DolphinAttack: Inaudible Voice Commands”, ACM CCS ’17. https://arxiv.org/abs/1708.09537
Evade model
𝐷𝐷𝑇𝑇𝑎𝑎𝑇𝑇𝐿𝐿𝑇𝑇𝑎𝑎ML
model Client
https://arxiv.org/abs/1705.07535https://arxiv.org/abs/1707.08945https://arxiv.org/abs/1708.09537
-
15
𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇
𝐿𝐿𝑇𝑇𝐿𝐿𝐿𝐿
Malicious client
Data owners
Analyst
ML model
Prediction Service Provider
API
Shokri et al., “Membership Inference Attacks Against Machine Learning Models”. IEEE S&P ’16. https://arxiv.org/pdf/1610.05820.pdfFredrikson et al., “Model Inversion Attacks that Exploit Confidence Information and Basic Countermeasures”. ACM CCS’15. https://www.cs.cmu.edu/~mfredrik/papers/fjr2015ccs.pdf
Invert model, infer membership
𝐷𝐷𝑇𝑇𝑎𝑎𝑇𝑇𝐿𝐿𝑇𝑇𝑎𝑎ML
model Client
Inference
𝐷𝐷𝑇𝑇𝑎𝑎𝑇𝑇𝐿𝐿𝑇𝑇𝑎𝑎
https://arxiv.org/pdf/1610.05820.pdfhttps://www.cs.cmu.edu/%7Emfredrik/papers/fjr2015ccs.pdf
-
16
𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇
𝐿𝐿𝑇𝑇𝐿𝐿𝐿𝐿
Malicious client
Data owners
Analyst
ML model
Prediction Service Provider
API Client
Tramer et al., “Stealing ML models via prediction APIs”, Usenix SEC ’16. https://arxiv.org/abs/1609.02943
Extract/steal model
𝐷𝐷𝑇𝑇𝑎𝑎𝑇𝑇𝐿𝐿𝑇𝑇𝑎𝑎ML
model
MLmodel
https://arxiv.org/abs/1609.02943
-
17
𝐿𝐿𝑇𝑇𝐿𝐿𝐿𝐿
Malicious prediction service
Data owners
Analyst
𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇 ML model
Prediction Service Provider
API Client X
Malmi and Weber. “You are what apps you use Demographic prediction based on user's apps”, ICWSM ‘16. https://arxiv.org/abs/1603.00059
Profile users
𝐷𝐷𝑇𝑇𝑎𝑎𝑇𝑇𝐿𝐿𝑇𝑇𝑎𝑎
𝐷𝐷𝑇𝑇𝑎𝑎𝑇𝑇𝐿𝐿𝑇𝑇𝐿𝐿𝑇𝑇Add: “X uses app”
Is this appmalicious?
https://arxiv.org/abs/1603.00059
-
18
Compromised toolchain: adversary inside training pipeline
Data owners
Analyst
𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇
𝐿𝐿𝑇𝑇𝐿𝐿𝐿𝐿
ML model
Prediction Service Provider
API Client
Song et al., “Machine Learning models that remember too much”, ACM CCS ’17. https://arxiv.org/abs/1709.07886 Hitja et al., “Deep Models Under the GAN: Information Leakage from Collaborative Deep Learning”, ACM CCS ’17. http://arxiv.org/abs/1702.07464
Sensitive query
Reveal trainingdata
Violate privacy
𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇
𝐿𝐿𝑇𝑇𝐿𝐿𝐿𝐿
ML model𝐷𝐷𝑇𝑇𝑎𝑎𝑇𝑇𝐿𝐿𝑇𝑇𝑎𝑎𝐷𝐷𝑇𝑇𝑎𝑎𝑇𝑇𝐿𝐿𝑇𝑇𝑎𝑎
https://arxiv.org/abs/1709.07886http://arxiv.org/abs/1702.07464
-
19
Malicious data owner
Data owners
Analyst
𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇𝑇
𝐿𝐿𝑇𝑇𝐿𝐿𝐿𝐿
ML model
Prediction Service Provider
API Client
https://www.theguardian.com/technology/2016/mar/26/microsoft-deeply-sorry-for-offensive-tweets-by-ai-chatbothttps://www.theguardian.com/technology/2017/nov/07/youtube-accused-violence-against-young-children-kids-content-google-pre-school-abuse
𝐷𝐷𝑇𝑇𝑎𝑎𝑇𝑇𝐿𝐿𝑇𝑇𝑎𝑎𝐷𝐷𝑇𝑇𝑎𝑎𝑇𝑇𝐿𝐿𝑇𝑇𝑎𝑎ML
model
Influence ML model (model poisoning)
https://www.theguardian.com/technology/2016/mar/26/microsoft-deeply-sorry-for-offensive-tweets-by-ai-chatbothttps://www.theguardian.com/technology/2017/nov/07/youtube-accused-violence-against-young-children-kids-content-google-pre-school-abuse
-
20
ML-based systems need to worry about adversarial behaviourAdversaries are multilateral, solutions are tooAdversarial machine learning is now a very active research area
Secure Systems Group Finnish Center for AIhttps://ssg.aalto.fi/ http://fcai.fi/
Summary
20
https://ssg.aalto.fi/http://fcai.fi/
-
Oblivious Neural NetworkPredictions via MiniONNTransformationsN. Asokan
http://asokan.org/asokan/@nasokan
(Joint work with Jian Liu, Mika Juuti, Yao Lu)By Source, Fair use, https://en.wikipedia.org/w/index.php?curid=54119040
https://en.wikipedia.org/w/index.php?curid=54119040
-
3
Machine learning as a service (MLaaS)
Predictions
Input
violation of clients’ privacy
-
4
Running predictions on client-side
Model
model theftevasionmodel inversion
-
Oblivious Neural Networks (ONN)
Given a neural network, is it possible to make it oblivious?
• server learns nothing about clients' input;
• clients learn nothing about the model.
5
-
6
Example: CryptoNets
FHE-encrypted input
FHE-encrypted predictions
[GDLLNW16] CryptoNets, ICML 2016
• High throughput for batch queries from same client • High overhead for single queries: 297.5s and 372MB (MNIST dataset)• Cannot support: high-degree polynomials, comparisons, …
FHE: Fully homomorphic encryption (https://en.wikipedia.org/wiki/Homomorphic_encryption)
http://proceedings.mlr.press/v48/gilad-bachrach16.pdfhttps://en.wikipedia.org/wiki/Homomorphic_encryption
-
7
MiniONN: Overview
Blinded input
Blinded predictions
oblivious protocols
• Low overhead: ~1s • Support all common neural networks
By Source, Fair use, https://en.wikipedia.org/w/index.php?curid=54119040
https://en.wikipedia.org/w/index.php?curid=54119040
-
8
Example
All operations are in a finite fieldx
y
'x
z
https://eprint.iacr.org/2017/452
Skip to performance
https://eprint.iacr.org/2017/452
-
9
Core idea: use secret sharing for oblivious computation
cy
cx'
cy' sy'+z
client & server have shares and s.t.
client & server have shares and s.t.
Use efficient cryptographic primitives (2PC, additively homomorphic encryption)
Skip to performance
-
10
Secret sharing initial input
Note that xc is independent of x. Can be pre-chosen
xSkip to performance
https://eprint.iacr.org/2017/452
https://eprint.iacr.org/2017/452
-
11
Compute locally by the server
Dot-product
Oblivious linear transformationSkip to performance
W•xc
-
12
Oblivious linear transformation: dot-productHomomorphic
Encryption with SIMD
u + v = W•xc; Note: u, v, and W•xc are independent of x. generated/stored in a precomputation phase
Skip to performance
-
13
Oblivious linear transformationSkip to performance
-
14
Oblivious linear transformationSkip to performance
-
Oblivious activation/pooling functions
15
Piecewise linear functions e.g.,• ReLU:• Oblivious ReLU:
- easily computed obliviously by a garbled circuit
Skip to performance
-
)1/(1: )(cs yycs exx +−+=+
Oblivious activation/pooling functions
16
Smooth functions e.g.,• Sigmoid:• Oblivious sigmoid:
- approximate by a piecewise linear function- then compute obliviously by a garbled circuit- empirically: ~14 segments sufficient
-
18
Combining the final result
They can jointly calculate max(y1,y2)(for minimizing information leakage)
-
19
Core idea: use secret sharing for oblivious computation
cy
cx'
cy' sy'+z
19
-
22
PTB/Sigmoid 4.39 (+ 13.9) 474 (+ 86.7) Less than 0.5%(cross-entropy loss)
Performance (for single queries)
Pre-computation phase timings in parentheses
CIFAR-10/ReLU 472 (+ 72) 6226 (+ 3046) none
Model Latency (s) Msg sizes (MB) Loss of accuracy
MNIST/Square 0.4 (+ 0.88) 44 (+ 3.6) none
PTB = Penn Treebank
-
MiniONN pros and cons
300-700x faster than CryptoNets
Can transform any given neural network to its oblivious variant
Still ~1000x slower than without privacy
Server can no longer filter requests or do sophisticated metering
Assumes online connectivity to server
Reveals structure (but not params) of NN
23
Skip to End
-
Can trusted computing help?
Hardware support for- Isolated execution: Trusted Execution Environment- Protected storage: Sealing- Ability to report status to a remote verifier:
Attestation
24
Other Software
Trusted Software
Protected Storage
Root of Trust
https://www.ibm.com/security/cryptocards/ https://www.infineon.com/tpm https://software.intel.com/en-us/sgxhttps://www.arm.com/products/security-on-arm/trustzone
Cryptocards Trusted Platform Modules ARM TrustZone Intel Software Guard Extensions
https://www.ibm.com/security/cryptocards/https://www.infineon.com/tpmhttps://software.intel.com/en-us/sgxhttps://www.arm.com/products/security-on-arm/trustzone
-
25
Using a client-side TEE to vet input
1. Attest client’s TEE app3. Input
4. Input, “Input/Metering Certificate”
5. MiniONN protocol + “Input/Metering Certificate”
2. Provision filtering policy
MiniONN + policy filtering + advanced metering
-
26
3. Input
Using a client-side TEE to run the model
1. Attest client’s TEE app
4. Predictions + “Metering Certificate”
2. Provision model configuration, filtering policy
MiniONN + policy filtering + advanced metering+ disconnected operation + performance + better privacy- harder to reason about model secrecy
5. “Metering Certificate”
-
27
2. Input
Using a server-side TEE to run the model
1. Attest server’s TEE app
3. Provision model configuration, filtering policy
MiniONN + policy filtering + advanced metering- disconnected operation + performance + better privacy
1. Attest server’s TEE app
4. Prediction
-
29
MiniONN: Efficiently transform any givenneural network into oblivious form with no/negligible accuracy loss
Trusted Computing can help realize improved security and privacy for ML
ML is very fragile in adversarial settingshttps://eprint.iacr.org/2017/452ACM CCS 2017
https://eprint.iacr.org/2017/452
ML-and-security.pdfMachine Learning�in the presence of adversariesMachine Learning is ubiquitousMachine Learning for security/privacyHow do we evaluate ML-based systems?Security & privacy of machine learningAdversarial examplesAdversarial examplesAdversarial examplesSlide Number 9Slide Number 10Model inversionModel inversionMachine Learning pipelineCompromised inputMalicious clientMalicious clientMalicious prediction serviceCompromised toolchain: �adversary inside training pipelineMalicious data ownerSummary
MiniONN.pdfOblivious Neural Network Predictions via MiniONN TransformationsMachine learning as a service (MLaaS)Running predictions on client-sideOblivious Neural Networks (ONN)Example: CryptoNetsMiniONN: OverviewExampleCore idea: use secret sharing for oblivious computationSecret sharing initial inputOblivious linear transformationOblivious linear transformation: dot-productOblivious linear transformationOblivious linear transformationOblivious activation/pooling functionsOblivious activation/pooling functionsCombining the final resultCore idea: use secret sharing for oblivious computationPerformance (for single queries)MiniONN pros and consCan trusted computing help?Using a client-side TEE to vet inputUsing a client-side TEE to run the modelUsing a server-side TEE to run the modelSlide Number 29