Deep Learning for Embedded Security Evaluation

Transcription

1 Deep Learning for Embedded Security Evaluation Emmanuel Prouff 1 1 Laboratoire de Sécurité des Composants, ANSSI, France April 2018, CISCO April 2018, CISCO E. Prouff 1/22

2 Contents 1. Context and Motivation 1.1 Illustration 1.2 Template Attacks 1.3 Countermeasures 2. A machine learning approach to classification 2.1 Introduction 2.2 The Underlying Classification Problem 2.3 Convolutional Neural Networks 2.4 Training of Models 3. Building a Community Around The Subject 3.1 ASCAD Open Data-Base 3.2 Leakage Characterization and Training 3.3 Results 4. Conclusions April 2018, CISCO E. Prouff 2/22

3 Probability distribution function (pdf) of Electromagnetic Emanations Cryptographic Processing with a secret k = 1. April 2018, CISCO E. Prouff 3/22

4 Probability distribution function (pdf) of Electromagnetic Emanations Cryptographic Processing with a secret k = 1. April 2018, CISCO E. Prouff 3/22

5 Probability distribution function (pdf) of Electromagnetic Emanations Cryptographic Processing with a secret k = 2. April 2018, CISCO E. Prouff 3/22

6 Probability distribution function (pdf) of Electromagnetic Emanations Cryptographic Processing with a secret k = 3. April 2018, CISCO E. Prouff 3/22

7 Probability distribution function (pdf) of Electromagnetic Emanations Cryptographic Processing with a secret k = 4. April 2018, CISCO E. Prouff 3/22

8 Context: Target Device Clone Device April 2018, CISCO E. Prouff 4/22

9 Context: Target Device Clone Device [On Clone Device] For every k estimate the pdf of X K = k. k = 1 k = 2 k = 3 k = 4 April 2018, CISCO E. Prouff 4/22

10 Context: Target Device Clone Device [On Clone Device] For every k estimate the pdf of X K = k. k = 1 k = 2 k = 3 k = 4 April 2018, CISCO E. Prouff 4/22

11 Context: Target Device Clone Device [On Clone Device] For every k estimate the pdf of X K = k. k = 1 k = 2 k = 3 [On Target Device] Estimate the pdf of X. k = 4 k =? April 2018, CISCO E. Prouff 4/22

12 Context: Target Device Clone Device [On Clone Device] For every k estimate the pdf of X K = k. k = 1 k = 2 k = 3 [On Target Device] Estimate the pdf of X. k = 4 k =? [Key-recovery] Compare the pdf estimations. April 2018, CISCO E. Prouff 4/22

13 Side Channel Attacks (Classical Approach) Notations X observation of the device behaviour (consumption, electro-magnetic emanation, timing, etc.) P public input of the processing Z target (a cryptographic sensitive variable Z = f (P, K)) Goal: make inference over Z, observing X April 2018, CISCO E. Prouff 5/22

14 Side Channel Attacks (Classical Approach) Notations X observation of the device behaviour (consumption, electro-magnetic emanation, timing, etc.) P public input of the processing Z target (a cryptographic sensitive variable Z = f (P, K)) Goal: make inference over Z, observing X Pr[Z X] April 2018, CISCO E. Prouff 5/22

15 Side Channel Attacks (Classical Approach) Notations X observation of the device behaviour (consumption, electro-magnetic emanation, timing, etc.) P public input of the processing Z target (a cryptographic sensitive variable Z = f (P, K)) Goal: make inference over Z, observing X Pr[Z X] Template Attacks Profiling phase (using profiling traces under known Z) Attack phase (N attack traces x i, e.g. with known plaintexts p i ) Log-likelihood score for each key hypothesis k N d k = log Pr[ X = x i Z = f (p i, k)] i=1 April 2018, CISCO E. Prouff 5/22

16 Side Channel Attacks (Classical Approach) Notations X observation of the device behaviour (consumption, electro-magnetic emanation, timing, etc.) P public input of the processing Z target (a cryptographic sensitive variable Z = f (P, K)) Goal: make inference over Z, observing X Pr[Z X] Template Attacks Profiling phase (using profiling traces under known Z) estimate Pr[ X Z = z] by simple distributions for each value of z Attack phase (N attack traces x i, e.g. with known plaintexts p i ) Log-likelihood score for each key hypothesis k N d k = log Pr[ X = x i Z = f (p i, k)] i=1 April 2018, CISCO E. Prouff 5/22

17 Side Channel Attacks (Classical Approach) Notations X observation of the device behaviour (consumption, electro-magnetic emanation, timing, etc.) P public input of the processing Z target (a cryptographic sensitive variable Z = f (P, K)) Goal: make inference over Z, observing X Pr[Z X] Template Attacks Profiling phase (using profiling traces under known Z) estimate Pr[ X Z = z] for each value of z Attack phase (N attack traces x i, e.g. with known plaintexts p i ) Log-likelihood score for each key hypothesis k N d k = log Pr[ X = x i Z = f (p i, k)] i=1 April 2018, CISCO E. Prouff 5/22

18 Side Channel Attacks (Classical Approach) Notations X observation of the device behaviour (consumption, electro-magnetic emanation, timing, etc.) P public input of the processing Z target (a cryptographic sensitive variable Z = f (P, K)) Goal: make inference over Z, observing X Pr[Z X] Template Attacks Profiling phase (using profiling traces under known Z) mandatory dimensionality reduction estimate Pr[ X Z = z] for each value of z Attack phase (N attack traces x i, e.g. with known plaintexts p i ) Log-likelihood score for each key hypothesis k N d k = log Pr[ X = x i Z = f (p i, k)] i=1 April 2018, CISCO E. Prouff 5/22

19 Side Channel Attacks (Classical Approach) Notations X observation of the device behaviour (consumption, electro-magnetic emanation, timing, etc.) P public input of the processing Z target (a cryptographic sensitive variable Z = f (P, K)) Goal: make inference over Z, observing X Pr[Z X] Template Attacks Profiling phase (using profiling traces under known Z) manage de-synchronization problem mandatory dimensionality reduction estimate Pr[ X Z = z] for each value of z Attack phase (N attack traces x i, e.g. with known plaintexts p i ) Log-likelihood score for each key hypothesis k N d k = log Pr[ X = x i Z = f (p i, k)] i=1 April 2018, CISCO E. Prouff 5/22

20 Defensive Mechanisms Misaligning Countermeasures Random Delays, Clock Jittering,... In theory: assume to be insufficient to provide security In practice: one of the main issues for evaluators = Need for efficient resynchronization techniques April 2018, CISCO E. Prouff 6/22

21 Defensive Mechanisms Misaligning Countermeasures Random Delays, Clock Jittering,... In theory: assume to be insufficient to provide security In practice: one of the main issues for evaluators = Need for efficient resynchronization techniques Masking Countermeasure Each key-dependent internal state element is randomly split into 2 shares The crypto algorithm is adapted to always manipulate shares at times The adversary needs to recover information on the two shares to recover K = Need for efficient Methods to recover tuple of leakage samples that jointly depend on the target secret April 2018, CISCO E. Prouff 6/22

22 Contents 1. Context and Motivation 1.1 Illustration 1.2 Template Attacks 1.3 Countermeasures 2. A machine learning approach to classification 2.1 Introduction 2.2 The Underlying Classification Problem 2.3 Convolutional Neural Networks 2.4 Training of Models 3. Building a Community Around The Subject 3.1 ASCAD Open Data-Base 3.2 Leakage Characterization and Training 3.3 Results 4. Conclusions April 2018, CISCO E. Prouff 7/22

23 Motivating Conclusions An evaluator can spend time to adjust realignment and/or to find points of interest... try realignments based over other kind of patterns modify parameters... April 2018, CISCO E. Prouff 8/22

24 Motivating Conclusions An evaluator can spend time to adjust realignment and/or to find points of interest... try realignments based over other kind of patterns modify parameters......but no prior knowledge about informative patterns no way to evaluate realignment without launching the afterwards attack April 2018, CISCO E. Prouff 8/22

25 Motivating Conclusions An evaluator can spend time to adjust realignment and/or to find points of interest... try realignments based over other kind of patterns modify parameters......but no prior knowledge about informative patterns no way to evaluate realignment without launching the afterwards attack Now: preprocessing to prepare data make strong hypotheses on the statistical dependency characterization to extract information The proposed perspective: preprocessing to prepare data make strong hypotheses on the statistical dependency Train algorithms to directly extract information April 2018, CISCO E. Prouff 8/22

26 Side Channel Attacks Notations X side channel trace Z target (a cryptographic sensitive variable Z = f (P, K)) Goal: make inference over Z, observing X Pr[Z X] Template Attacks Machine Learning Side Channel Attacks Profiling phase (using profiling traces under known Z) manage de-synchronization problem mandatory dimensionality reduction estimate Pr[ X Z = z] for each value of z Attack phase (N attack traces, e.g. with known plaintexts p i ) Log-likelihood score for each key hypothesis k d k = N log Pr[ X = x i Z = f (p i, k)] i=1 April 2018, CISCO E. Prouff 9/22

27 Side Channel Attacks with a Classifier Notations X side channel trace Z target (a cryptographic sensitive variable Z = f (P, K)) Goal: make inference over Z, observing X Pr[Z X] Template Attacks Machine Learning Side Channel Attacks Profiling phase (using profiling traces under known Z) manage de-synchronization problem mandatory dimensionality reduction estimate Pr[ X Z = z] for each value of z Attack phase (N attack traces, e.g. with known plaintexts p i ) Log-likelihood score for each key hypothesis k d k = N log Pr[ X = x i Z = f (p i, k)] i=1 April 2018, CISCO E. Prouff 9/22

28 Side Channel Attacks with a Classifier Notations X side channel trace Z target (a cryptographic sensitive variable Z = f (P, K)) Goal: make inference over Z, observing X Pr[Z X] Template Attacks Machine Learning Side Channel Attacks Training phase (using training traces under known Z) manage de-synchronization problem mandatory dimensionality reduction estimate Pr[ X Z = z] for each value of z Attack phase (N attack traces, e.g. with known plaintexts p i ) Log-likelihood score for each key hypothesis k d k = N log Pr[ X = x i Z = f (p i, k)] i=1 April 2018, CISCO E. Prouff 9/22

29 Side Channel Attacks with a Classifier Notations X side channel trace Z target (a cryptographic sensitive variable Z = f (P, K)) Goal: make inference over Z, observing X Pr[Z X] Template Attacks Machine Learning Side Channel Attacks Training phase (using training traces under known Z) manage de-synchronization problem mandatory dimensionality reduction construct a classifier F ( x) = y Pr[Z X = x] Attack phase (N attack traces, e.g. with known plaintexts p i ) Log-likelihood score for each key hypothesis k d k = N log Pr[ X = x i Z = f (p i, k)] i=1 April 2018, CISCO E. Prouff 9/22

30 Side Channel Attacks with a Classifier Notations X side channel trace Z target (a cryptographic sensitive variable Z = f (P, K)) Goal: make inference over Z, observing X Pr[Z X] Template Attacks Machine Learning Side Channel Attacks Training phase (using training traces under known Z) manage de-synchronization problem mandatory dimensionality reduction construct a classifier F ( x) = y Pr[Z X = x] Attack phase (N attack traces, e.g. with known plaintexts p i ) Log-likelihood score for each key hypothesis k d k = N log F ( x i )[f (p i, k)] i=1 April 2018, CISCO E. Prouff 9/22

31 Side Channel Attacks with a Classifier Notations X side channel trace Z target (a cryptographic sensitive variable Z = f (P, K)) Goal: make inference over Z, observing X Pr[Z X] Template Attacks Machine Learning Side Channel Attacks Training phase (using training traces under known Z) } manage de-synchronization problem mandatory dimensionality reduction Integrated approach construct a classifier F ( x) = y Pr[Z X = x] Attack phase (N attack traces, e.g. with known plaintexts p i ) Log-likelihood score for each key hypothesis k d k = N log F ( x i )[f (p i, k)] i=1 April 2018, CISCO E. Prouff 9/22

32 Classification Classification problem Assign to a datum X (e.g. an image) a label Z among a set of possible labels Z = {Cat, Dog, Horse} Classifier April 2018, CISCO E. Prouff 10/22

33 Classification Classification problem Assign to a datum X (e.g. an image) a label Z among a set of possible labels Z = {Cat, Dog, Horse} Classifier Classification 0% 20% 40% 60% Horse Dog Cat April 2018, CISCO E. Prouff 10/22

34 Classification Classification problem Assign to a datum X (e.g. an image) a label Z among a set of possible labels Z = {Cat, Dog, Horse} Classifier Classification 0% 20% 40% 60% Horse Dog Cat Pr[Z X] April 2018, CISCO E. Prouff 10/22

35 Classification Classification problem Assign to a datum X (e.g. an image) a label Z among a set of possible labels Z = {Cat, Dog, Horse} Classifier Classification 0% 20% 40% 60% Horse Dog Cat Pr[Z X] SCA as a Classification Problem x Classifier P(Z X=x) 0% 50% 100% Z=1 Z=0 April 2018, CISCO E. Prouff 10/22

36 Machine Learning Approach Overview of Machine Learning Methodology Human effort: choose a class of algorithms choose a model to fit + tune hyper-parameters Automatic training: automatic tuning of trainable parameters to fit data aaa April 2018, CISCO E. Prouff 11/22

37 Machine Learning Approach Overview of Machine Learning Methodology Human effort: choose a class of algorithms Neural Networks choose a model to fit + tune hyper-parameters Automatic training: automatic tuning of trainable parameters to fit data Stochastic Gradient Descent aaa April 2018, CISCO E. Prouff 11/22

38 Machine Learning Approach Overview of Machine Learning Methodology Human effort: choose a class of algorithms Neural Networks choose a model to fit + tune hyper-parameters MLP, ConvNet Automatic training: automatic tuning of trainable parameters to fit data Stochastic Gradient Descent aaa April 2018, CISCO E. Prouff 11/22

39 Machine Learning Approach Overview of Machine Learning Methodology Human effort: choose a class of algorithms Neural Networks choose a model to fit + tune hyper-parameters MLP, ConvNet Automatic training: automatic tuning of trainable parameters to fit data Stochastic Gradient Descent aaa April 2018, CISCO E. Prouff 11/22

40 Convolutional Neural Networks An answer to translation-invariance Classifier Classification 0% 20% 40% 60% Horse Dog Cat April 2018, CISCO E. Prouff 12/22

41 Convolutional Neural Networks An answer to translation-invariance Classifier Classification 0% 20% 40% 60% Horse Dog Cat April 2018, CISCO E. Prouff 12/22

42 Convolutional Neural Networks An answer to translation-invariance Classifier Classification 0% 20% 40% 60% Horse Dog Cat April 2018, CISCO E. Prouff 12/22

43 Convolutional Neural Networks An answer to translation-invariance Classifier Classification 0% 20% 40% 60% Horse Dog Cat It is important to explicit the data translation-invariance April 2018, CISCO E. Prouff 12/22

44 Convolutional Neural Networks An answer to translation-invariance?? Classification? Classifier 0% 10% 20% 30% 40% 50% Horse Dog Cat It is important to explicit the data translation-invariance April 2018, CISCO E. Prouff 12/22

45 Convolutional Neural Networks An answer to translation-invariance x Classifier P(Z X=x) 0% 50% 100% Z=1 Z=0 It is important to explicit the data translation-invariance April 2018, CISCO E. Prouff 12/22

46 Convolutional Neural Networks An answer to translation-invariance x Classifier P(Z X=x) 0% 50% 100% Z=1 Z=0 It is important to explicit the data translation-invariance April 2018, CISCO E. Prouff 12/22

47 Convolutional Neural Networks An answer to translation-invariance x Classifier P(Z X=x) 0% 50% 100% Z=1 Z=0 It is important to explicit the data translation-invariance April 2018, CISCO E. Prouff 12/22

48 Convolutional Neural Networks An answer to translation-invariance x Classifier P(Z X=x) 0% 50% 100% Z=1 Z=0 It is important to explicit the data translation-invariance Convolutional Neural Networks: share weights across space April 2018, CISCO E. Prouff 12/22

49 Training of Neural Networks Trading Side-Channel Expertise for Deep Learning Expertise... or huge computational power! Training Aims at finding the parameters of the model (aka approximation) for the the statistical dependency between the target value and the leakage. The approximation is done by solving a minimization problem with respect to some metric (aka cost function) The training algorithm has itself some training hyper-parameters: the number of iterations (aka epochs) of the minimization procedure, the number of input traces (aka batch) treated during a single iteration. The architecture of the trained model has architecture hyper-parameters: the size of the layers, the nature of the layers, the number of layers, etc. Finding sound hyper-parameters is the main issue in Deep Learning: this can be done thanks to a good understanding of the underlying structure of the data and/or access to important computational power. April 2018, CISCO E. Prouff 13/22

50 ConvNet typical architecture 2nd Block of CONV layers n conv = 3 and n filters,2 = 8 3rd Block of CONV layers n conv = 3 and n filters,3 = 16 input trace l = 700 1st Block of CONV layers n conv = 3 and n filters,1 = 4 Max POOL (w = 2) Max POOL (w = 2) Block of FC layers + SOFTMAX n dense = 3 April 2018, CISCO E. Prouff 14/22

51 Contents 1. Context and Motivation 1.1 Illustration 1.2 Template Attacks 1.3 Countermeasures 2. A machine learning approach to classification 2.1 Introduction 2.2 The Underlying Classification Problem 2.3 Convolutional Neural Networks 2.4 Training of Models 3. Building a Community Around The Subject 3.1 ASCAD Open Data-Base 3.2 Leakage Characterization and Training 3.3 Results 4. Conclusions April 2018, CISCO E. Prouff 15/22

52 Creation of an open database for Training and Testing ANSSI recently publishes source codes of secure cryptographic implementations for public 8-bit architectures ( data-bases of electromagnetic leakages ( example scripts for the training and testing of Deep Learning models in the context of security evaluations Goal Enable easy fair and easy benchmarking Initiate discussions and exchanges on the application of DL to SCA Create a community of contributors on this subject April 2018, CISCO E. Prouff 16/22

53 Nature of the Observations/Traces Side-channel observations in ASCAD correspond to the masked processing of a simple cryptographic primitive Information leakage validated thanks to SNR characterization snr2 snr Time Samples April 2018, CISCO E. Prouff 17/22

54 Nature of the Observations/Traces Side-channel observations in ASCAD correspond to the masked processing of a simple cryptographic primitive Information leakage validated thanks to SNR characterization snr2 snr Time Samples Validate that shares are manipulated at different times April 2018, CISCO E. Prouff 17/22

55 Nature of the Observations/Traces Side-channel observations in ASCAD correspond to the masked processing of a simple cryptographic primitive Information leakage validated thanks to SNR characterization snr2 snr Time Samples Validate that shares are manipulated at different times Scripts are also proposed to add artificial signal jittering April 2018, CISCO E. Prouff 17/22

56 Our Training Strategy Find a base model architecture and find training hyper-parameters for which a convergence towards the good key hypothesis is visible Fine-tune all the hyper-parameters one after another to get the best efficiency/effectiveness trade-off Table: Benchmarks Summary Parameter Reference Metric Range Choice Training Parameters Epochs - rank vs time 10, 25, 50, 60,..., 100, 150 up to 100 Batch Size - rank vs time 50, 100, Architecture Parameters Blocks n blocks rank, accuracy [2..5] 5 CONV layers n conv rank, accuracy [0..3] 1 Filters n filters,1 rank vs time {2 i ; i [4..7]} 64 Kernel Size - rank {3, 6, 11} 11 FC Layers n dense rank, accuracy vs time [0..3] 2 ACT Function α rank ReLU, Sigmoid, Tanh ReLU Pooling Layer - rank Max, Average, Stride Average Padding - rank Same, Valid Same April 2018, CISCO E. Prouff 18/22

57 The Base Architecture Mean rank of the good-key hypothesis obtained with VGG-16, ResNet-50 and Inception-v3 w.r.t. different epochs: April 2018, CISCO E. Prouff 19/22

58 Comparisons with State-Of-the-Art Methods April 2018, CISCO E. Prouff 20/22

59 Contents 1. Context and Motivation 1.1 Illustration 1.2 Template Attacks 1.3 Countermeasures 2. A machine learning approach to classification 2.1 Introduction 2.2 The Underlying Classification Problem 2.3 Convolutional Neural Networks 2.4 Training of Models 3. Building a Community Around The Subject 3.1 ASCAD Open Data-Base 3.2 Leakage Characterization and Training 3.3 Results 4. Conclusions April 2018, CISCO E. Prouff 21/22

60 Conclusions State-of-the-Art Template Attack separates resynchronization/dimensionality reduction from characterization April 2018, CISCO E. Prouff 22/22

61 Conclusions State-of-the-Art Template Attack separates resynchronization/dimensionality reduction from characterization Deep Learning provides an integrated approach to directly extract information from rough data (no preprocessing) April 2018, CISCO E. Prouff 22/22

62 Conclusions State-of-the-Art Template Attack separates resynchronization/dimensionality reduction from characterization Deep Learning provides an integrated approach to directly extract information from rough data (no preprocessing) Many recent results validate the practical interest of the Machine Learning approach April 2018, CISCO E. Prouff 22/22

63 Conclusions State-of-the-Art Template Attack separates resynchronization/dimensionality reduction from characterization Deep Learning provides an integrated approach to directly extract information from rough data (no preprocessing) Many recent results validate the practical interest of the Machine Learning approach We are in the very beginning and we are still discovering how much Deep Learning is efficient April 2018, CISCO E. Prouff 22/22

64 Conclusions State-of-the-Art Template Attack separates resynchronization/dimensionality reduction from characterization Deep Learning provides an integrated approach to directly extract information from rough data (no preprocessing) Many recent results validate the practical interest of the Machine Learning approach We are in the very beginning and we are still discovering how much Deep Learning is efficient New needs: big data-bases for the training, platforms to enable comparisons and benchmarking, create an open community ML for Embedded Security Analysis, encourage exchanges with the Machine Learning community, understand the efficiency of the current countermeasures April 2018, CISCO E. Prouff 22/22

65 Conclusions State-of-the-Art Template Attack separates resynchronization/dimensionality reduction from characterization Deep Learning provides an integrated approach to directly extract information from rough data (no preprocessing) Many recent results validate the practical interest of the Machine Learning approach We are in the very beginning and we are still discovering how much Deep Learning is efficient New needs: big data-bases for the training, platforms to enable comparisons and benchmarking, create an open community ML for Embedded Security Analysis, encourage exchanges with the Machine Learning community, understand the efficiency of the current countermeasures Thank You! Questions? April 2018, CISCO E. Prouff 22/22