PrivApprox. Privacy- Preserving Stream Analytics.

Transcription

1 PrivApprox Privacy- Preserving Stream Analytics Do Le Quoc, Martin Beck, Pramod Bhatotia, Ruichuan Chen, Christof Fetzer, Thorsten Strufe July 2017

2 Motivation Clients Analysts Private data Recommendation, Ads 1

3 Motivation Clients Analysts Private data Recommendation, Ads Strong privacy guarantee 1

4 Motivation Clients Analysts Private data Recommendation, Ads Strong privacy guarantee High utility analytics in real- time 1

5 Motivation Clients Analysts Private data Recommendation, Ads Strong privacy guarantee High utility analytics in real- time How to preserve users privacy while supporting high- utility data analytics for low- latency stream processing? 1

6 State- of- the- art systems Clients 2

7 State- of- the- art systems Clients Personal data should be stored locally under the clients control 2

8 State- of- the- art systems Clients Personal data should be stored locally under the clients control 2

9 State- of- the- art systems Clients 2

10 State- of- the- art systems Clients Aggregator Analyst 2

11 State- of- the- art systems Clients Forward query Aggregator Submit query Analyst 2

12 State- of- the- art systems Clients Aggregator Analyst 2

13 State- of- the- art systems Clients Answer query Aggregator Analyst 2

14 State- of- the- art systems Clients Answer query Aggregator Add noise Analyst 2

15 State- of- the- art systems Clients Answer query Aggregator Add noise Privacy- preserving output Analyst 2

16 State- of- the- art systems Clients Answer query Aggregator Add noise Privacy- preserving output Analyst Differential Privacy 2

17 State- of- the- art systems Clients Answer query Aggregator Add noise Privacy- preserving output Analyst Differential Privacy Limitations: 2

18 State- of- the- art systems Clients Answer query Aggregator Add noise Privacy- preserving output Analyst Differential Privacy Limitations: Deal with only single- shot batch queries L 2

19 State- of- the- art systems Clients Answer query Aggregator Add noise Privacy- preserving output Analyst Differential Privacy Limitations: Deal with only single- shot batch queries L Require synchronization between system components L 2

20 State- of- the- art systems Clients Answer query Aggregator Add noise Privacy- preserving output Analyst Differential Privacy Limitations: Deal with only single- shot batch queries L Require synchronization between system components L Require a trusted aggregator L 2

21 PrivApprox Clients PrivApprox Analyst 3

22 PrivApprox Clients PrivApprox Analyst PrivApprox: 3

23 PrivApprox Clients PrivApprox Analyst PrivApprox: Supports stream processing with low latency J 3

24 PrivApprox Clients PrivApprox Analyst PrivApprox: Supports stream processing with low latency J Enables a truly synchronization- free distributed architecture J 3

25 PrivApprox Clients PrivApprox Analyst PrivApprox: Supports stream processing with low latency J Enables a truly synchronization- free distributed architecture J Requires lower trust in aggregator J 3

26 Outline Motivation Overview Design Evaluation 4

27 System overview Clients PrivApprox Analyst 5

28 System overview Clients PrivApprox (Query, budget) Analyst 5

29 System overview Clients PrivApprox (Query, budget) Analyst Execution budget: Latency/throughput guarantees Desired computing resources for query processing Desired accuracy 5

30 System overview Clients PrivApprox (Query, budget) Analyst Execution budget: Latency/throughput guarantees Desired computing resources for query processing Desired accuracy 5

31 System overview Clients PrivApprox Analyst 5

32 System overview Clients PrivApprox Result Analyst 5

33 System overview Clients PrivApprox Approximate computing Result Analyst 5

34 System overview Clients PrivApprox Approximate computing Result Analyst Low latency 5

35 System overview Clients PrivApprox Approximate computing + Randomized response Result Analyst Low latency 5

36 System overview Clients PrivApprox Approximate computing + Randomized response Result Analyst Low latency Privacy 5

37 System overview Clients PrivApprox Approximate computing + Randomized response Result Analyst 5

38 System overview Clients PrivApprox Approximate computing + Randomized response Result Analyst Zero- knowledge Privacy 5

39 System overview Clients PrivApprox Approximate computing + Randomized response Result Analyst Zero- knowledge Privacy Zero- knowledge Privacy > Differential Privacy 5

40 #1: Approximate computing 6

41 #1: Approximate computing State- of- the- art- systems Compute Add noise (Privacy- preserving) approximate output 6

42 #1: Approximate computing State- of- the- art- systems Compute Add noise (Privacy- preserving) approximate output Idea: To achieve low latency, compute over a sub- set of data items instead of the entire data- set 6

43 #1: Approximate computing State- of- the- art- systems Compute Add noise (Privacy- preserving) approximate output Idea: To achieve low latency, compute over a sub- set of data items instead of the entire data- set Take a sample Approximate computing Compute Approximate output ± Error bound 6

44 #2: Randomized response 7

45 #2: Randomized response Idea: To preserve privacy, clients may not need to provide truthful answers every time 7

46 #2: Randomized response Idea: To preserve privacy, clients may not need to provide truthful answers every time Client 7

47 #2: Randomized response Idea: To preserve privacy, clients may not need to provide truthful answers every time Client 7

48 #2: Randomized response Idea: To preserve privacy, clients may not need to provide truthful answers every time Client Truthful Answer 7

49 #2: Randomized response Idea: To preserve privacy, clients may not need to provide truthful answers every time Client Truthful Answer 7

50 #2: Randomized response Idea: To preserve privacy, clients may not need to provide truthful answers every time Client No Truthful Answer Yes 7

51 #2: Randomized response Idea: To preserve privacy, clients may not need to provide truthful answers every time Client No Truthful Answer Yes Provides plausible deniability for clients responding to sensitive queries; achieves differential privacy (RAPPOR [CCS 14]) 7

52 Outline Motivation Overview Design Evaluation 8

53 Query model 9

54 Query model Divide answer s value range into buckets, enforce a binary answer in each bucket 9

55 Query model Divide answer s value range into buckets, enforce a binary answer in each bucket Query: SELECT age FROM clients WHERE city = Santa Clara 9

56 Query model Divide answer s value range into buckets, enforce a binary answer in each bucket Query: SELECT age FROM clients WHERE city = Santa Clara >60 9

57 Query model Divide answer s value range into buckets, enforce a binary answer in each bucket Query: SELECT age FROM clients WHERE city = Santa Clara Age: >60 9

58 Query model Divide answer s value range into buckets, enforce a binary answer in each bucket Query: SELECT age FROM clients WHERE city = Santa Clara Age: >60 Client cannot arbitrarily manipulate answers 9

59 Workflow: Submit query Aggregator (Query, budget) Analyst 10

60 Workflow: Submit query Clients Aggregator (Query, budget) Analyst Cost- Function(budget) System parameters: Sampling parameter Randomized response parameters 10

61 Workflow: Submit query Clients (Query, parameters) Aggregator (Query, budget) Analyst Cost- Function(budget) System parameters: Sampling parameter Randomized response parameters 10

62 Workflow: Answer query 11

63 Workflow: Answer query Client 11

64 Workflow: Answer query Client Step #1 Sampling (Flip a coin to decide to answer query or not) 11

65 Workflow: Answer query Client Step #1 Step #2 Sampling (Flip a coin to decide to answer query or not) Randomized Response 11

66 Workflow: Answer query Client Step #1 Step #2 Step #3 Sampling (Flip a coin to decide to answer query or not) Randomized Response Send randomized answer 11

67 Workflow: Answer query Client Step #1 Step #2 Step #3 Sampling (Flip a coin to decide to answer query or not) Randomized Response Send randomized answer Zero- knowledge privacy 11

68 Workflow: Answer query Client Step #1 Step #2 Step #3 Sampling (Flip a coin to decide to answer query or not) Randomized Response Send randomized answer Zero- knowledge privacy See the paper for details! 11

69 Workflow: Answer query Clients Randomized answers Aggregator 12

70 Workflow: Answer query Clients Randomized answers Aggregator Approximate result ± Error bound Analyst 12

71 Workflow: Answer query Clients Randomized answers Aggregator Approximate result ± Error bound Analyst Lack of anonymity and unlinkability? 12

72 #3: Anonymity and unlinkability 13

73 #3: Anonymity and unlinkability Idea: XOR- based Encryption 13

74 #3: Anonymity and unlinkability Idea: XOR- based Encryption Client 13

75 #3: Anonymity and unlinkability Idea: XOR- based Encryption Client Encrypt answer M: GenerateKey - > M k M XOR M k - > M E 13

76 #3: Anonymity and unlinkability Idea: XOR- based Encryption Client Proxy Aggregator Proxy Encrypt answer M: GenerateKey - > M k M XOR M k - > M E 13

77 #3: Anonymity and unlinkability Idea: XOR- based Encryption Client Proxy Aggregator Proxy Encrypt answer M: GenerateKey - > M k M XOR M k - > M E Decrypt answer M E : M E XOR M k - > M 13

78 Implementation Clients Proxy Aggregator Analyst Proxy 14

79 Implementation Clients Proxy Aggregator Analyst Proxy 14

80 Implementation Clients Proxy Aggregator Analyst Proxy 14

81 Implementation Clients Proxy Aggregator Analyst Proxy 14

82 Outline Motivation Overview Design Evaluation 15

83 Experimental setup Evaluation questions Utility vs privacy Throughput & latency Network overhead 16

84 Experimental setup Evaluation questions Utility vs privacy Throughput & latency Network overhead See the paper for more results! 16

85 Experimental setup Evaluation questions Utility vs privacy Throughput & latency Network overhead See the paper for more results! Testbed Cluster: 44 nodes Dataset: NYC Taxi ride records, household electricity usage 16

86 Accuracy vs privacy 17

87 Accuracy vs privacy 0.6 Randomization parameters #1 (p = 0.6, q = 0.6) Randomization parameters #2 (p = 0.9, q = 0.6) 6 Accuracy loss (%) Privacy (ε zk ) Sampling Fraction (%) 0 17

88 Accuracy vs privacy 0.6 Randomization parameters #1 (p = 0.6, q = 0.6) Randomization parameters #2 (p = 0.9, q = 0.6) 6 The lower the better Accuracy loss (%) Accuracy loss Privacy level Privacy (ε zk ) Sampling Fraction (%) Trade- off between utility and privacy 0 17

89 Accuracy vs privacy 0.6 Randomization parameters #1 (p = 0.6, q = 0.6) Randomization parameters #2 (p = 0.9, q = 0.6) 6 The lower the better Accuracy loss (%) Accuracy loss Privacy level Privacy (ε zk ) Sampling Fraction (%) Trade- off between utility and privacy 0 17

90 Accuracy vs privacy 0.6 Randomization parameters #1 (p = 0.6, q = 0.6) Randomization parameters #2 (p = 0.9, q = 0.6) 6 The lower the better Accuracy loss (%) Accuracy loss Privacy level Privacy (ε zk ) Sampling Fraction (%) Trade- off between utility and privacy 0 17

91 Accuracy vs privacy 0.6 Randomization parameters #1 (p = 0.6, q = 0.6) Randomization parameters #2 (p = 0.9, q = 0.6) 6 The lower the better Accuracy loss (%) Accuracy loss Privacy level Privacy (ε zk ) Sampling Fraction (%) Trade- off between utility and privacy 0 17

92 Accuracy vs privacy 0.6 Randomization parameters #1 (p = 0.6, q = 0.6) Randomization parameters #2 (p = 0.9, q = 0.6) 6 The lower the better Accuracy loss (%) Accuracy loss Privacy level Privacy (ε zk ) Sampling Fraction (%) Trade- off between utility and privacy 0 17

93 Throughput 18

94 Throughput NYC Taxi Ride Household Electricity Throughput (K) #nodes 18

95 Throughput NYC Taxi Ride Household Electricity The higher the better Throughput (K) #nodes 18

96 Throughput NYC Taxi Ride Household Electricity The higher the better Throughput (K) #nodes 18

97 Throughput NYC Taxi Ride Household Electricity The higher the better Throughput (K) #nodes ~8X speedup when going from one node to 20 nodes 18

98 Latency 19

99 Latency NYC Taxi Ride Household Electricity Total processing time (seconds) Native Sampling fraction (%) 19

100 Latency NYC Taxi Ride Household Electricity The lower the better Total processing time (seconds) Native Sampling fraction (%) 19

101 Latency NYC Taxi Ride Household Electricity The lower the better Total processing time (seconds) Native Sampling fraction (%) 19

102 Latency NYC Taxi Ride Household Electricity The lower the better Total processing time (seconds) Native Sampling fraction (%) ~1.66X lower than the native execution with sampling fraction of 60% 19

103 Network overhead 20

104 Network overhead NYC Taxi Ride Household Electricity Network traffic (GB) Native Sampling fraction (%) 20

105 Network overhead NYC Taxi Ride Household Electricity The lower the better Network traffic (GB) Native Sampling fraction (%) 20

106 Network overhead NYC Taxi Ride Household Electricity The lower the better Network traffic (GB) Native Sampling fraction (%) 20

107 Network overhead NYC Taxi Ride Household Electricity The lower the better Network traffic (GB) Native Sampling fraction (%) ~1.6X lower than the native execution with sampling fraction of 60% 20

108 Conclusion PrivApprox: a privacy- preserving stream analytics system over distributed datasets 21

109 Conclusion PrivApprox: a privacy- preserving stream analytics system over distributed datasets Privacy Zero- knowledge privacy 21

110 Conclusion PrivApprox: a privacy- preserving stream analytics system over distributed datasets Privacy Practical Zero- knowledge privacy Adaptive execution based on query budget 21

111 Conclusion PrivApprox: a privacy- preserving stream analytics system over distributed datasets Privacy Practical Efficient Zero- knowledge privacy Adaptive execution based on query budget Randomized response & sampling techniques 21

112 Conclusion PrivApprox: a privacy- preserving stream analytics system over distributed datasets Privacy Practical Efficient Zero- knowledge privacy Adaptive execution based on query budget Randomized response & sampling techniques Thank you! 21