Causative Adversarial Learning

Causative Adversarial Learning

Huang Xiao, am 24.06.2015xiaohu(at)in.tum.de

Talk presented on Deep Learning in Action @Munich

Motivation

Deep networks can be easily fooled … [1]

Evolution Algor. generated images

99.99% confidence

“It turns out some DNNs only focus on discriminative features in images.”

[1] Nguyen A, Yosinski J, Clune J. Deep Neural Networks are Easily Fooled: High Confidence Predictions for Unrecognizable Images. In Computer Vision and Pattern Recognition (CVPR '15), IEEE, 2015.

Motivation

Spam alerts

Google brain, 16000 CPUs

Learning is expensive!

Motivation

Adversarial Learning

Reverse engineering of machine learning. It aims to design robust and secure learning

algorithms.

Big Picture

Are the modern learning systems really secure?

Training dataset Model

Test (Validation)

dataset

Training Test

Update

● Increase test error

● Reduce learning accuracy

● Fool the intelligent system

● Achieve personal gain

Big Picture

Are the modern learning systems really secure?

Training dataset Model

Test (Validation)

dataset

Training Test

Update

Causative Attack

Exploratory Attack

Attack’s capability

Access to Data Knowledge about features

Knowledge about the classifier

Limited Knowledge Partially Maybe Yes

Perfect Knowledge Yes Yes Yes

These are real inputs from users.

Basics

❏ Observations❏ True signal:❏ Polynomial curve fitting ❏ is unknown❏ => learn the green curve

Observation

Original signal

Least square

TrainingMinimize empirical squared error.

Estimated output

Observed output

Least square

TrainingMinimize empirical squared error.

OverfittingEstimated output

Observed output

Overfitting

❏ Bad on unseen test set❏ Central problem of ML.❏ Generalization❏ E.g., regularization, prior,

more data, model selection

Bias-Variance

❏ Trade off❏ Overfitting == low bias, high variance❏ Underfitting == high bias, low variance❏ Noise is dominating!

W is very sensitive

Bias Variance Decomposition

Objective

Increase bias or variance?

Types of Adversaries

● Causative Attack (Poisoning)○ Understanding how the learning algorithms work○ Engineering on features or labels of training set○ Change the discriminant function

● Exploratory Attack (Evasion)○ Engineering features of a test point○ Circumvent the legitimate detection○ Change the discriminant result

Types of Adversaries

● Causative Attack (Poisoning)○ Understanding how the learning algorithms work○ Engineering on features or labels of training set○ Change the discriminant function

● Exploratory Attack (Evasion)○ Engineering features of a test point○ Circumvent the legitimate detection○ Change the discriminant result

Label Noises on SVM

● SVM: One of the state-of-art classifier● Binary case: +1, -1● Label flips attack under a certain budget● Maximizing error on validation set● Methods:

○ ALFA○ Distance based: far-first, near-first, random○ Continuous relaxation gradient ascend○ Correlated cluster

BasicsWe measure the error on a validation set using the function trained on training set.

A training data set

A validation data set

Classifier trained on

Regularization coefficient

Risk measurement on validation set

Flip Labels

Flip Labels

Huang Xiao, B. Biggio, B. Nelson, Han Xiao, C. Eckert, and F. Roli, “Support Vector Machines under Adversarial Label Contamination”, Neurocomputing, vol. Special Issue on Advances in Learning with Label Noise, In Press.

Poisoning Attack on SVM

● Noises on features, not on labels● Design a malicious training point● Maximizing the error (e.g., test error,

hinge loss, ...)● Gradient ascend

How to?

Retrain the SVM after injecting a malicious point ,, , move the point such that the classification error on validation set is maximized.

Validation data set with m samples

SVM trained on training set with a malicious point

Poisoning Attack on SVM

Poisoning Attack on SVM

B. Biggio, B. Nelson, and P. Laskov, “Poisoning attacks against support vector machines”, in 29th Int'l Conf. on Machine Learning (ICML), 2012.

Walking example

B. Biggio, B. Nelson, and P. Laskov, “Poisoning attacks against support vector machines”, in 29th Int'l Conf. on Machine Learning (ICML), 2012

You can:● Mimic the ‘9’ as ‘8’ or,● Label a ‘9’ as a ‘8’

Poisoning Lasso

● Lasso: feature selection, more generally, L1 regularization

● Feature selection is often the first step for many learning system

● Other targets: Rigid regression, elastic network

● Gradient based method

Lasso

Capture the most relevant features in data set automatically by shrinking the feature weights.

from:Tibshirani, R. (1996). Regression shrinkage and selection via the lasso. J. Royal. Statist. Soc B., Vol. 58, No. 1, pages 267-288).

Feature selection

x1

x2

x3

x4

x5

x6

x7

x8

x9

x10

5.1 4.6 4.5 4.0 4.0 1.8 0 0 0 0

Non-zero (weight) features are selected for next stage training!

Feature selection

x1

x2

x3

x4

x5

x6

x7

x8

x9

x10

5.13.64.23.14.21.80000

Non-zero (weight) features are selected for next stage training!

Adding a malicious point

Training set

Intuition

# features

# S

amp

les

# features

# S

amp

les

#samples ≪ #features #samples ≫ #features

Intuition

# features

# S

amp

les

# features

# S

amp

les

#samples ≪ #features #samples ≫ #features

Danger!

Add some random noises

Research goals

● Investigating robustness of feature selection algorithms

● Design a multiple point attack method ● Warning: feature selection might not be

reliable● A gradient based poisoning framework

Objective function

We inject a malicious point to form a new compromised Data .

Variable: , we are maximising w.r.t

Remark that is learnt on contaminated data .

Maximise Generalization Error!

Gradient Ascent

Update rule:

descent

ascent

min

max

bound box

Demonstration

Error surface

Initial attack point

on each (x, y)

Xiao, Huang, Battista Biggio, Gavin Brown, Giorgio Fumera, Claudia Eckert, and Fabio Roli. Is Feature Selection Secure against Training Data Poisoning?. In ICML'15,Lille, France, July 2015.

Demonstration

Gradient ascend path

Xiao, Huang, Battista Biggio, Gavin Brown, Giorgio Fumera, Claudia Eckert, and Fabio Roli. Is Feature Selection Secure against Training Data Poisoning?. In ICML'15,Lille, France, July 2015.

Wrap up

● Don’t expect your algorithms too fancy● Don’t expect adversaries too silly ● Setup objective and do the worst-case

study● Machine learning needs to be more

robust● There’s no innocent data

Thank you, question?