Alleviating Privacy Attacks via
Causal Learning
Shruti Tople, Amit Sharma, Aditya V. Nori
Microsoft Research
https://arxiv.org/abs/1909.12732
https://github.com/microsoft/robustdg
Motivation: ML models leak information
about data points in the training set
Neural
Network
TrainingHealth Records
(HIV/AIDS
patients)
ML-as-a-service
Member of
Train Dataset
Non-member
Membership Inference Attacks
[SP’17][CSF’18][NDSS’19][SP’19]
The likely reason is overfitting
Output
85%
Output
95%
Overfitting to
dataset
• Neural networks or associational models
overfit to the training dataset
• Membership inference adversary exploits
differences in prediction score for training and
test data [CSF’18]
Overfitting to
distribution
The likely reason is overfitting
• Neural networks or associational models
overfit to the training dataset
• Membership inference attacks exploit
differences in prediction score for training and
test data [CSF’18]
• Privacy risk can increase when model is
deployed to different distributions
• E.g., Hospital in one region shares the model to
other regions
Output
85%
Output
95%
Overfitting to
dataset
Output
75%
Poor generalization across distributions exacerbates
membership inference risk.
Can causal ML
models help?
Can causal ML models help?
Contributions
1. Causal models provide stronger (differential) privacy guarantees than
associational models.
• Due to their better generalizability on new distributions.
2. And hence are more robust to membership inference attacks.
• As the training dataset size → ∞, membership inference attack’s accuracy drops to a
random guess.
3. We empirically demonstrate privacy benefits of causal models across 5 datasets.
• Associational models exhibit up to 80% attack accuracy whereas causal models exhibit
attack accuracy close to 50%.
Causal
Learning
Privacy
Disease
Severity
Background: Causal Learning
𝒀
Blood
Pressure
Heart
Rate
𝑿 𝒑𝒂𝒓𝒆𝒏𝒕 𝑿 𝒑𝒂𝒓𝒆𝒏𝒕
𝑿 𝟏 𝑿 𝟐
Weight Age
Use a structural causal model (SCM) that defines what
conditional probabilities are invariant across different
distributions [Pearl’09].
Background: Causal Learning
Use a structural causal model (SCM) that defines what
conditional probabilities are invariant across different
distributions [Pearl’09].
Causal Predictive Model: A prediction model based only
on the parents of the outcome Y.
What if SCM is not known? Learn an invariant feature
representation across distributions [ABGD’19, MTS’20].
For ML models, causal learning can be useful for
fairness [KLRS’17]
explainability [DSZ’16, MTS’19]
privacy [this work]
Disease
Severity
𝒀
Blood
Pressure
Heart
Rate
𝑿 𝒑𝒂𝒓𝒆𝒏𝒕 𝑿 𝒑𝒂𝒓𝒆𝒏𝒕
𝑿 𝟏 𝑿 𝟐
Weight Age
𝒀
𝑋𝑆0 𝑋 𝑃𝐴
𝑋𝑆2
𝑋𝑆1
𝑋 𝐶𝐻
𝑋𝑐𝑝
Intervention
Why is a model based on causal parents
invariant across data distributions?
Why is a model based on causal parents
invariant across data distributions?
𝒀
𝑋𝑆0 𝑋 𝑃𝐴
𝑋𝑆2
𝑋𝑆1
𝑋 𝐶𝐻
𝑋𝑐𝑝
Intervention
𝒀
𝑋𝑆0 𝑋 𝑃𝐴
𝑋𝑆2
𝑋𝑆1
𝑋 𝐶𝐻
𝑋𝑐𝑝
𝑃(𝑌|𝑋 𝑃𝐴) is invariant across different distributions, unless there is a
change in true data-generating process for Y.
Result 1: Worst-case out-of-distribution error of a
causal model is lower than an associational model.
For any model ℎ, and 𝑃∗ such that 𝑃∗ 𝑌 𝑋 𝑃𝐴 = 𝑃(𝑌|𝑋 𝑃𝐴),
In-Distribution Error (IDE)= 𝐈𝐃𝐄 𝐏 𝒉, 𝒚 = 𝐋 𝑷 𝒉, 𝒚 − 𝐋 𝑺∼P(𝒉, 𝒚)
Expected loss on the same distribution as the train data
Out-of-Distribution Error (ODE)=𝐎𝐃𝐄 𝐏,𝐏∗ 𝒉, 𝒚 = 𝐋 𝑷∗ 𝒉, 𝒚 − 𝐋 𝑺∼P 𝒉, 𝒚
Expected loss on a different distribution 𝑃∗
than the train data
Result 1: Worst-case out-of-distribution error of a
causal model is lower than an associational model.
For any model ℎ, and 𝑃∗ such that 𝑃∗ 𝑌 𝑋 𝑃𝐴 = 𝑃(𝑌|𝑋 𝑃𝐴),
In-Distribution Error (IDE)= 𝐈𝐃𝐄 𝐏 𝒉, 𝒚 = 𝐋 𝑷 𝒉, 𝒚 − 𝐋 𝑺∼P(𝒉, 𝒚)
Expected loss on the same distribution as the train data
Out-of-Distribution Error (ODE)=𝐎𝐃𝐄 𝐏,𝐏∗ 𝒉, 𝒚 = 𝐋 𝑷∗ 𝒉, 𝒚 − 𝐋 𝑺∼P 𝒉, 𝒚
Expected loss on a different distribution 𝑃∗
than the train data
Proof Idea. Simple case: Assume 𝑦 = 𝑓(𝒙) is deterministic.
𝐎𝐃𝐄 𝐏,𝐏∗ 𝒉 𝐜, 𝒚 ≤ 𝐈𝐃𝐄 𝐏(𝒉 𝒄, 𝒚) + 𝒅𝒊𝒔𝒄 𝐋 𝑷, 𝑷∗
Discrepancy
b/w 𝑷 and 𝑷∗
distributions
Causal Model
Result 1: Worst-case out-of-distribution error of a
causal model is lower than an associational model.
For any model ℎ, and 𝑃∗ such that 𝑃∗ 𝑌 𝑋 𝑃𝐴 = 𝑃(𝑌|𝑋 𝑃𝐴),
In-Distribution Error (IDE)= 𝐈𝐃𝐄 𝐏 𝒉, 𝒚 = 𝐋 𝑷 𝒉, 𝒚 − 𝐋 𝑺∼P(𝒉, 𝒚)
Expected loss on the same distribution as the train data
Out-of-Distribution Error (ODE)=𝐎𝐃𝐄 𝐏,𝐏∗ 𝒉, 𝒚 = 𝐋 𝑷∗ 𝒉, 𝒚 − 𝐋 𝑺∼P 𝒉, 𝒚
Expected loss on a different distribution 𝑃∗
than the train data
Proof Idea. Simple case: Assume 𝑦 = 𝑓(𝒙) is deterministic.
𝐎𝐃𝐄 𝐏,𝐏∗ 𝒉 𝐜, 𝒚 ≤ 𝐈𝐃𝐄 𝐏(𝒉 𝒄, 𝒚) + 𝒅𝒊𝒔𝒄 𝐋 𝑷, 𝑷∗
𝐎𝐃𝐄 𝐏,𝐏∗ 𝒉 𝒂, 𝒚 ≤ 𝐈𝐃𝐄 𝐏 𝒉 𝒂, 𝒚 + 𝒅𝒊𝒔𝒄 𝐋 𝑷, 𝑷∗
+ 𝐋 𝑷∗(𝒉 𝒂,𝑷
𝑶𝑷𝑻
, 𝒚)
⇒ max
𝐏∗
𝐎𝐃𝐄𝐁𝐨𝐮𝐧𝐝 𝐏,𝐏∗ 𝒉 𝐜, 𝒚 ≤ max
𝐏∗
𝐎𝐃𝐄𝐁𝐨𝐮𝐧𝐝 𝐏,𝐏∗ 𝒉 𝒂, 𝒚
Discrepancy
b/w 𝑷 and 𝑷∗
distributions
Optimal 𝒉 𝒂 on P is
not optimal on 𝑷∗
Causal Model
Assoc. Model
Result 1: Worst-case out-of-distribution error of a
causal model is lower than an associational model.
And better generalization results in lower
sensitivity for a causal model
Sensitivity: If a single data point 𝒙, 𝑦 ∼ 𝑃∗ is added to the train
dataset 𝑆 to create 𝑆′, how much does the learnt model h 𝑆
min
change?
Since the optimal causal model is the same across all 𝑃∗
, adding
any 𝒙, 𝑦 ∼ 𝑃∗ has less impact on a trained causal model.
Sensitivity for a causal
model
Sensitivity for an
associational model
Main Result: A causal model has stronger
Differential Privacy guarantees
Let M be a mechanism that returns a ML model trained over dataset 𝑆, M(𝑆) = ℎ.
Differential Privacy [DR’14]: A learning mechanism M satisfies 𝜖-differential
privacy if for any two datasets, 𝑆, 𝑆′ that differ in one data point,
Pr(M 𝑆 ∈𝐻)
Pr(M 𝑆′ ∈𝐻)
≤ 𝑒 𝜖.
(Smaller 𝜖 values provide better privacy guarantees)
Since lower sensitivity ⇒ lower 𝜖,
Theorem: When equivalent Laplace noise is added and models are trained on same
dataset, causal mechanism MC provides 𝜖 𝐶-DP and associational mechanism MA
provides 𝜖 𝐴-DP guarantees such that:
𝝐 𝒄 ≤ 𝝐 𝑨
Therefore, causal models are more robust to
membership inference (MI) attacks
Advantage of an MI adversary:
(True Positive Rate – False Positive Rate)
in detecting whether 𝑥 is from training dataset or not.
[From Yeom et al. CSF’18] Membership advantage of an adversary is bounded by
𝑒 𝜖
− 1.
Since the optimal causal models are the same for 𝑃 and 𝑃∗,
As 𝑛 → ∞, membership advantage of causal model → 0.
Theorem: When trained on the same dataset of size 𝑛, membership
advantage of a causal model is lower than the membership advantage for an
associational model.
Empirical
Evaluation
Goal: Compare MI attack accuracy between
causal and associational models
[BN] When true causal structure is known
Datasets generated from Bayesian networks: Child, Sachs, Water, Alarm
Causal model: MLE estimation based on Y’s parents
Associational model: Neural networks with 3 linear layers
𝑃∗: Noise added to conditional probabilities (uniform or additive)
[MNIST] When true causal structure is unknown
Colored MNIST dataset (Digits are correlated with color)
Causal Model: Invariant Risk Minimization that utilizes 𝑃 𝑌 𝑋 𝑃𝐴 is same across distributions [ABGD’19]
Associational Model: Empirical Risk Minimization using the same NN architecture
𝑃∗: Different correlations between color and digit than the train dataset
Attacker Model: Predict whether an input belongs to train dataset or not
[BN] With uniform noise, MI attack accuracy
for a causal model is near a random guess
80%
50%
For associational models, the attacker can guess membership in training set with 80% accuracy.
[BN-Child] With uniform noise, MI attack accuracy
for a causal model is near a random guess
80%
50%
For associational models, the attacker can guess membership in training set with 80% accuracy.
Privacy without loss in utility: Causal & DNN models achieve same prediction accuracy.
[BN-Child] MI Attack accuracy increases with
amount of noise for associational models, but
stays constant at 50% for causal models
[BN] Consistent results across all four datasets
High attack accuracy for associational
models when 𝑃∗
(Test2) has uniform noise.
Same classification accuracy between
causal and associational models.
[MNIST] MI attack accuracy is lower for invariant
risk minimizer compared to associational model
IRM model motivated by causal reasoning has 53% attack accuracy, close to random.
Associational model also fails to generalize: 16% accuracy on test set.
Model
Train
Accuracy
(%)
Test
Accuracy
(%)
Attack
Accuracy
(%)
Causal Model
(IRM)
70 69 53
Associational
Model (ERM)
87 16 66
Conclusion
• Established theoretical connection between causality and differential privacy.
• Demonstrated the benefits of causal ML models for alleviating privacy attacks,
both theoretically and empirically.
• Code available at https://github.com/microsoft/robustdg
Future work: Investigate robustness of causal models with other kinds of
adversarial attacks.
Causal
Learning
Privacy
thank you!
Amit Sharma
Microsoft Research
References
• [ABGD’19] Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv
preprint arXiv:1907.02893, 2019.
• [CSF’18] Yeom, S., Giacomelli, I., Fredrikson, M., and Jha, S. Privacy risk in machine learning: Analyzing the connection
to overfitting. CSF 2018.
• [DR’14] Cynthia Dwork, Aaron Roth, et al. The algorithmic foundations of differential privacy. Foundations and
Trends in Theoretical Computer Science, 9(3–4):211–407, 2014.
• [DSZ’16] Anupam Datta, Shayak Sen, and Yair Zick. Algorithmic transparency via quantitative input influence: Theory
and experiments with learning systems. In Security and Privacy (SP), 2016 IEEE Symposium on, pp. 598–617. IEEE,
2016
• [KLRS’17] Matt J Kusner, Joshua Loftus, Chris Russell, and Ricardo Silva. Counterfactual fairness. In Advances in
Neural Information Processing Systems, pp. 4066–4076, 2017.
• [MTS’19] Mahajan, Divyat, Chenhao Tan, and Amit Sharma. "Preserving Causal Constraints in Counterfactual
Explanations for Machine Learning Classifiers." arXiv preprint arXiv:1912.03277 (2019).
• [MTS’20] Mahajan, Divyat, Shruti Tople and Amit Sharma. “Domain Generalization using Causal Matching”. arXiv
preprint arXiv:2006.07500, 2020.
• [NDSS’19] Salem, A., Zhang, Y., Humbert, M., Fritz, M., and Backes, M. Ml-leaks: Model and data independent
membership inference attacks and defenses on machine learning models. NDSS 2019.
• [SP’17] Shokri, R., Stronati, M., Song, C., and Shmatikov, V. Membership inference attacks against machine learning
models. Security and Privacy (SP), 2017.
• [SP’19] Nasr, M., Shokri, R., and Houmansadr, A. Comprehensive privacy analysis of deep learning: Stand-alone and
federated learning under passive and active white-box inference attacks. Security and Privacy (SP), 2019.

Alleviating Privacy Attacks Using Causal Models

  • 1.
    Alleviating Privacy Attacksvia Causal Learning Shruti Tople, Amit Sharma, Aditya V. Nori Microsoft Research https://arxiv.org/abs/1909.12732 https://github.com/microsoft/robustdg
  • 2.
    Motivation: ML modelsleak information about data points in the training set Neural Network TrainingHealth Records (HIV/AIDS patients) ML-as-a-service Member of Train Dataset Non-member Membership Inference Attacks [SP’17][CSF’18][NDSS’19][SP’19]
  • 3.
    The likely reasonis overfitting Output 85% Output 95% Overfitting to dataset • Neural networks or associational models overfit to the training dataset • Membership inference adversary exploits differences in prediction score for training and test data [CSF’18]
  • 4.
    Overfitting to distribution The likelyreason is overfitting • Neural networks or associational models overfit to the training dataset • Membership inference attacks exploit differences in prediction score for training and test data [CSF’18] • Privacy risk can increase when model is deployed to different distributions • E.g., Hospital in one region shares the model to other regions Output 85% Output 95% Overfitting to dataset Output 75% Poor generalization across distributions exacerbates membership inference risk.
  • 5.
  • 6.
    Can causal MLmodels help? Contributions 1. Causal models provide stronger (differential) privacy guarantees than associational models. • Due to their better generalizability on new distributions. 2. And hence are more robust to membership inference attacks. • As the training dataset size → ∞, membership inference attack’s accuracy drops to a random guess. 3. We empirically demonstrate privacy benefits of causal models across 5 datasets. • Associational models exhibit up to 80% attack accuracy whereas causal models exhibit attack accuracy close to 50%. Causal Learning Privacy
  • 7.
    Disease Severity Background: Causal Learning 𝒀 Blood Pressure Heart Rate 𝑿𝒑𝒂𝒓𝒆𝒏𝒕 𝑿 𝒑𝒂𝒓𝒆𝒏𝒕 𝑿 𝟏 𝑿 𝟐 Weight Age Use a structural causal model (SCM) that defines what conditional probabilities are invariant across different distributions [Pearl’09].
  • 8.
    Background: Causal Learning Usea structural causal model (SCM) that defines what conditional probabilities are invariant across different distributions [Pearl’09]. Causal Predictive Model: A prediction model based only on the parents of the outcome Y. What if SCM is not known? Learn an invariant feature representation across distributions [ABGD’19, MTS’20]. For ML models, causal learning can be useful for fairness [KLRS’17] explainability [DSZ’16, MTS’19] privacy [this work] Disease Severity 𝒀 Blood Pressure Heart Rate 𝑿 𝒑𝒂𝒓𝒆𝒏𝒕 𝑿 𝒑𝒂𝒓𝒆𝒏𝒕 𝑿 𝟏 𝑿 𝟐 Weight Age
  • 9.
    𝒀 𝑋𝑆0 𝑋 𝑃𝐴 𝑋𝑆2 𝑋𝑆1 𝑋𝐶𝐻 𝑋𝑐𝑝 Intervention Why is a model based on causal parents invariant across data distributions?
  • 10.
    Why is amodel based on causal parents invariant across data distributions? 𝒀 𝑋𝑆0 𝑋 𝑃𝐴 𝑋𝑆2 𝑋𝑆1 𝑋 𝐶𝐻 𝑋𝑐𝑝 Intervention 𝒀 𝑋𝑆0 𝑋 𝑃𝐴 𝑋𝑆2 𝑋𝑆1 𝑋 𝐶𝐻 𝑋𝑐𝑝 𝑃(𝑌|𝑋 𝑃𝐴) is invariant across different distributions, unless there is a change in true data-generating process for Y.
  • 11.
    Result 1: Worst-caseout-of-distribution error of a causal model is lower than an associational model.
  • 12.
    For any modelℎ, and 𝑃∗ such that 𝑃∗ 𝑌 𝑋 𝑃𝐴 = 𝑃(𝑌|𝑋 𝑃𝐴), In-Distribution Error (IDE)= 𝐈𝐃𝐄 𝐏 𝒉, 𝒚 = 𝐋 𝑷 𝒉, 𝒚 − 𝐋 𝑺∼P(𝒉, 𝒚) Expected loss on the same distribution as the train data Out-of-Distribution Error (ODE)=𝐎𝐃𝐄 𝐏,𝐏∗ 𝒉, 𝒚 = 𝐋 𝑷∗ 𝒉, 𝒚 − 𝐋 𝑺∼P 𝒉, 𝒚 Expected loss on a different distribution 𝑃∗ than the train data Result 1: Worst-case out-of-distribution error of a causal model is lower than an associational model.
  • 13.
    For any modelℎ, and 𝑃∗ such that 𝑃∗ 𝑌 𝑋 𝑃𝐴 = 𝑃(𝑌|𝑋 𝑃𝐴), In-Distribution Error (IDE)= 𝐈𝐃𝐄 𝐏 𝒉, 𝒚 = 𝐋 𝑷 𝒉, 𝒚 − 𝐋 𝑺∼P(𝒉, 𝒚) Expected loss on the same distribution as the train data Out-of-Distribution Error (ODE)=𝐎𝐃𝐄 𝐏,𝐏∗ 𝒉, 𝒚 = 𝐋 𝑷∗ 𝒉, 𝒚 − 𝐋 𝑺∼P 𝒉, 𝒚 Expected loss on a different distribution 𝑃∗ than the train data Proof Idea. Simple case: Assume 𝑦 = 𝑓(𝒙) is deterministic. 𝐎𝐃𝐄 𝐏,𝐏∗ 𝒉 𝐜, 𝒚 ≤ 𝐈𝐃𝐄 𝐏(𝒉 𝒄, 𝒚) + 𝒅𝒊𝒔𝒄 𝐋 𝑷, 𝑷∗ Discrepancy b/w 𝑷 and 𝑷∗ distributions Causal Model Result 1: Worst-case out-of-distribution error of a causal model is lower than an associational model.
  • 14.
    For any modelℎ, and 𝑃∗ such that 𝑃∗ 𝑌 𝑋 𝑃𝐴 = 𝑃(𝑌|𝑋 𝑃𝐴), In-Distribution Error (IDE)= 𝐈𝐃𝐄 𝐏 𝒉, 𝒚 = 𝐋 𝑷 𝒉, 𝒚 − 𝐋 𝑺∼P(𝒉, 𝒚) Expected loss on the same distribution as the train data Out-of-Distribution Error (ODE)=𝐎𝐃𝐄 𝐏,𝐏∗ 𝒉, 𝒚 = 𝐋 𝑷∗ 𝒉, 𝒚 − 𝐋 𝑺∼P 𝒉, 𝒚 Expected loss on a different distribution 𝑃∗ than the train data Proof Idea. Simple case: Assume 𝑦 = 𝑓(𝒙) is deterministic. 𝐎𝐃𝐄 𝐏,𝐏∗ 𝒉 𝐜, 𝒚 ≤ 𝐈𝐃𝐄 𝐏(𝒉 𝒄, 𝒚) + 𝒅𝒊𝒔𝒄 𝐋 𝑷, 𝑷∗ 𝐎𝐃𝐄 𝐏,𝐏∗ 𝒉 𝒂, 𝒚 ≤ 𝐈𝐃𝐄 𝐏 𝒉 𝒂, 𝒚 + 𝒅𝒊𝒔𝒄 𝐋 𝑷, 𝑷∗ + 𝐋 𝑷∗(𝒉 𝒂,𝑷 𝑶𝑷𝑻 , 𝒚) ⇒ max 𝐏∗ 𝐎𝐃𝐄𝐁𝐨𝐮𝐧𝐝 𝐏,𝐏∗ 𝒉 𝐜, 𝒚 ≤ max 𝐏∗ 𝐎𝐃𝐄𝐁𝐨𝐮𝐧𝐝 𝐏,𝐏∗ 𝒉 𝒂, 𝒚 Discrepancy b/w 𝑷 and 𝑷∗ distributions Optimal 𝒉 𝒂 on P is not optimal on 𝑷∗ Causal Model Assoc. Model Result 1: Worst-case out-of-distribution error of a causal model is lower than an associational model.
  • 15.
    And better generalizationresults in lower sensitivity for a causal model Sensitivity: If a single data point 𝒙, 𝑦 ∼ 𝑃∗ is added to the train dataset 𝑆 to create 𝑆′, how much does the learnt model h 𝑆 min change? Since the optimal causal model is the same across all 𝑃∗ , adding any 𝒙, 𝑦 ∼ 𝑃∗ has less impact on a trained causal model. Sensitivity for a causal model Sensitivity for an associational model
  • 16.
    Main Result: Acausal model has stronger Differential Privacy guarantees Let M be a mechanism that returns a ML model trained over dataset 𝑆, M(𝑆) = ℎ. Differential Privacy [DR’14]: A learning mechanism M satisfies 𝜖-differential privacy if for any two datasets, 𝑆, 𝑆′ that differ in one data point, Pr(M 𝑆 ∈𝐻) Pr(M 𝑆′ ∈𝐻) ≤ 𝑒 𝜖. (Smaller 𝜖 values provide better privacy guarantees) Since lower sensitivity ⇒ lower 𝜖, Theorem: When equivalent Laplace noise is added and models are trained on same dataset, causal mechanism MC provides 𝜖 𝐶-DP and associational mechanism MA provides 𝜖 𝐴-DP guarantees such that: 𝝐 𝒄 ≤ 𝝐 𝑨
  • 17.
    Therefore, causal modelsare more robust to membership inference (MI) attacks Advantage of an MI adversary: (True Positive Rate – False Positive Rate) in detecting whether 𝑥 is from training dataset or not. [From Yeom et al. CSF’18] Membership advantage of an adversary is bounded by 𝑒 𝜖 − 1. Since the optimal causal models are the same for 𝑃 and 𝑃∗, As 𝑛 → ∞, membership advantage of causal model → 0. Theorem: When trained on the same dataset of size 𝑛, membership advantage of a causal model is lower than the membership advantage for an associational model.
  • 18.
  • 19.
    Goal: Compare MIattack accuracy between causal and associational models [BN] When true causal structure is known Datasets generated from Bayesian networks: Child, Sachs, Water, Alarm Causal model: MLE estimation based on Y’s parents Associational model: Neural networks with 3 linear layers 𝑃∗: Noise added to conditional probabilities (uniform or additive) [MNIST] When true causal structure is unknown Colored MNIST dataset (Digits are correlated with color) Causal Model: Invariant Risk Minimization that utilizes 𝑃 𝑌 𝑋 𝑃𝐴 is same across distributions [ABGD’19] Associational Model: Empirical Risk Minimization using the same NN architecture 𝑃∗: Different correlations between color and digit than the train dataset Attacker Model: Predict whether an input belongs to train dataset or not
  • 20.
    [BN] With uniformnoise, MI attack accuracy for a causal model is near a random guess 80% 50% For associational models, the attacker can guess membership in training set with 80% accuracy.
  • 21.
    [BN-Child] With uniformnoise, MI attack accuracy for a causal model is near a random guess 80% 50% For associational models, the attacker can guess membership in training set with 80% accuracy. Privacy without loss in utility: Causal & DNN models achieve same prediction accuracy.
  • 22.
    [BN-Child] MI Attackaccuracy increases with amount of noise for associational models, but stays constant at 50% for causal models
  • 23.
    [BN] Consistent resultsacross all four datasets High attack accuracy for associational models when 𝑃∗ (Test2) has uniform noise. Same classification accuracy between causal and associational models.
  • 24.
    [MNIST] MI attackaccuracy is lower for invariant risk minimizer compared to associational model IRM model motivated by causal reasoning has 53% attack accuracy, close to random. Associational model also fails to generalize: 16% accuracy on test set. Model Train Accuracy (%) Test Accuracy (%) Attack Accuracy (%) Causal Model (IRM) 70 69 53 Associational Model (ERM) 87 16 66
  • 25.
    Conclusion • Established theoreticalconnection between causality and differential privacy. • Demonstrated the benefits of causal ML models for alleviating privacy attacks, both theoretically and empirically. • Code available at https://github.com/microsoft/robustdg Future work: Investigate robustness of causal models with other kinds of adversarial attacks. Causal Learning Privacy thank you! Amit Sharma Microsoft Research
  • 26.
    References • [ABGD’19] MartinArjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019. • [CSF’18] Yeom, S., Giacomelli, I., Fredrikson, M., and Jha, S. Privacy risk in machine learning: Analyzing the connection to overfitting. CSF 2018. • [DR’14] Cynthia Dwork, Aaron Roth, et al. The algorithmic foundations of differential privacy. Foundations and Trends in Theoretical Computer Science, 9(3–4):211–407, 2014. • [DSZ’16] Anupam Datta, Shayak Sen, and Yair Zick. Algorithmic transparency via quantitative input influence: Theory and experiments with learning systems. In Security and Privacy (SP), 2016 IEEE Symposium on, pp. 598–617. IEEE, 2016 • [KLRS’17] Matt J Kusner, Joshua Loftus, Chris Russell, and Ricardo Silva. Counterfactual fairness. In Advances in Neural Information Processing Systems, pp. 4066–4076, 2017. • [MTS’19] Mahajan, Divyat, Chenhao Tan, and Amit Sharma. "Preserving Causal Constraints in Counterfactual Explanations for Machine Learning Classifiers." arXiv preprint arXiv:1912.03277 (2019). • [MTS’20] Mahajan, Divyat, Shruti Tople and Amit Sharma. “Domain Generalization using Causal Matching”. arXiv preprint arXiv:2006.07500, 2020. • [NDSS’19] Salem, A., Zhang, Y., Humbert, M., Fritz, M., and Backes, M. Ml-leaks: Model and data independent membership inference attacks and defenses on machine learning models. NDSS 2019. • [SP’17] Shokri, R., Stronati, M., Song, C., and Shmatikov, V. Membership inference attacks against machine learning models. Security and Privacy (SP), 2017. • [SP’19] Nasr, M., Shokri, R., and Houmansadr, A. Comprehensive privacy analysis of deep learning: Stand-alone and federated learning under passive and active white-box inference attacks. Security and Privacy (SP), 2019.