A Privacy Preserving Data Mining Methodology for Dynamically Predicting Emerging Human Threats

Transcription

1 DETC A Privacy Preserving Data Mining Methodology for Dynamically Predicting Emerging Human Threats Tuesday, August 6 th, 2013 Gautam Manohar & Conrad S. Tucker {gautam.atulya@gmail.com, ctucker4@psu.edu, } Introduction Manohar, Tucker

2 Presentation Overview Research Motivation and Background Methodology The Knowledge Discovery process Data Acquisition and Storage Data Mining Predictive Model Construction Result Interpretation and Output Application Case Study Results and Discussion Conclusion and Path Forward Presentation Overview Manohar, Tucker

3 RESEARCH MOTIVATION Research Motivation Manohar, Tucker

4 Motivation Research Motivation Manohar, Tucker

5 Tracking sample Tracking video Capturing Emergence Manohar, Tucker

6 Motivation and Background Existing systems are passive and more useful for post-incident analysis. Privacy issues with most existing systems become a hindrance in public use (I.e. the need to preserve Personally Identifiable Information (PII)) Research Motivation Manohar, Tucker

7 Why Individual Body Movement Data? BODY LANGUAGE "The most important thing in communication is to hear what isn't beingsaid." Peter F. Drucker Literature Review Manohar, Tucker

8 RESEARCH METHODOLOGY Research Methodology Manohar, Tucker

9 Proposed Methodology Research Methodology Manohar, Tucker

10 Step 1: Data Acquisition Data acquisition hardware setup consists of a sensor system with: an RGB video camera, and an infrared depth sensor Output from sensors is used to create a virtual skeleton of the subject with 20 nodes as shown Each nodes collects data pertaining to: 3D Spatial Coordinates (X,Y,Z) Timestamp Velocities of each node Research Methodology High Fidelity Data, Privacy Preserving Manohar, Tucker

11 Large Scale Data Base Research Methodology Manohar, Tucker

12 Proposed Methodology Research Methodology Manohar, Tucker

13 Step 2: Data Transfer and Storage The data is stored in a structured Relational Database with fields for the following measures: Timestamp Euclidean Coordinates Velocities of each node Boolean Threat Class defining whether the data collected during training was for a threat action or not. Research Methodology Manohar, Tucker

14 Step 2: Data Transfer and Storage The data is stored in a structured Relational Database with fields for the following measures: Research Methodology Manohar, Tucker

15 Proposed Methodology Research Methodology Manohar, Tucker

16 Step3: Data Mining/Knowledge Discovery Research Methodology Manohar, Tucker

17 Knowledge Discovery in Data Bases Supervised Learning Unsupervised Learning Research Methodology Manohar, Tucker

18 Supervised VS Unsupervised Learning Supervised y=f(x): true function D: labeled training set D: {x i,f(x i )} Learn: G(x): model trained to predict labels D Goal: E[(F(x)-G(x)) 2 ] 0 Well defined criteria: Accuracy, RMSE,... Unsupervised Generator: true model D: unlabeled data sample D: {x i } Learn Underlying data structure Goal: Find natural patterns Well defined criteria: varies Research Methodology Manohar, Tucker

19 Capturing Threat Emergence Time t 1 Time t n Time t n+1 Model(t 1 ) Model(t n ) Model(t n+1 ) Research Methodology Manohar, Tucker

20 Data Mining Decision Tree Induction Given a time stamped Data Set (t), Feature 1 Feature 2 Feature N Class A 1,1 A 2,1 A N,1 C j, A 1,M A 2,M A N,M C j,m Entropy( T ) p( C T )log p( C T ) = j T GAIN( X ) Entropy( T ) Entropy ( T ) Gain ratio(x) = j k i = X i i= 1 T k i= 1 2 Gain( X ) Ti Ti log2 T T j Tucker C., H.M. Kim,"Trend Mining for Predictive Product Design", Transactions of ASME: Journal of Mechanical Design, Vol. 133, No. 11, Research Methodology Manohar, Tucker

21 Features Time Series Gain Ratio Predict t1 t2 t3 t4 t5 t6 t7 t8 t9 t10 t11 t12 t13_predict X_Elbow Joint Y_Hip_Joint X_Shoulder X_Accel_Arm Y_Accel_Hip Z_Arm_Joint Feature Gain Ratio Plot Over Time Gain Ratio Research Methodology Time Hard Drive TalkTime Camera Interface Connectivity Manohar, Tucker X_Elbow Joint Y_Hip_Joint X_Shoulder X_Accel_Arm Y_Accel_Hip Z_Arm_Joint 2 G Processor 21

22 Features Time Series Gain Ratio Predict t1 t2 t3 t4 t5 t6 t7 t8 t9 t10 t11 t12 t13_predict X_Elbow Joint Y_Hip_Joint X_Shoulder X_Accel_Arm Y_Accel_Hip Z_Arm_Joint Feature Gain Ratio Plot Over Time Gain Ratio Research Methodology Time X_Elbow Joint Hard Drive Y_Hip_Joint TalkTime X_Shoulder Camera X_Accel_Arm Interface Y_Accel_Hip Connectivity Z_Arm_Joint Manohar, Tucker G Processor 22

23 n- time stamped data sets No IM(Feature (i), Data Set (t)) i=i+1 Data set (t)=n No Predict IM(Feature (i)) Yes i=i+1 Split Data Sets 1,,n based on Max Predicted IM (Feature(1), Feature (k)) For Each Subset, P (Class 1) Yes End TREE, Classify Irrelevant Features Manohar, Tucker

24 n- time stamped data sets No IM(Feature (i), Data Set (t)) i=i+1 No Data set (t)=n Predict IM(Feature (i)) Yes i=i+1 Split Data Sets 1,,n based on Max Predicted IM (Feature(1), Feature (k)) For Each Subset, P (Class 1) Yes End TREE, Classify Irrelevant Features Manohar, Tucker

25 Holt-Winters Forecasting The (k) step-ahead forecasting model is defined as: y ( k ) = L + kt + I t t t s + k t Where: Level Lt (the level component): L = α( y I ) + (1 α)( L + T ) t t t s t 1 t 1 Trend Tt (the slope component): T = γ( L L ) + (1 γ) T t t t 1 t 1 Season It (the seasonal component): I = δ( y L) + (1 δ) I t t t t s The smoothing parameters α,γ δ, are in the range {0,1} Research Methodology Manohar, Tucker

26 Features Time Series Gain Ratio Predict t1 t2 t3 t4 t5 t6 t7 t8 t9 t10 t11 t12 t13_predict X_Elbow Joint Y_Hip_Joint X_Shoulder X_Accel_Arm Y_Accel_Hip Z_Arm_Joint Feature Gain Ratio Plot Over Time Gain Ratio Research Methodology Time Hard Drive X_Elbow Joint TalkTime Y_Hip_Joint Camera X_Shoulder Interface X_Accel_Arm Manohar, Tucker Connectivity Y_Accel_Hip 2 Z_Arm_Joint G Processor

27 n- time stamped data sets No IM(Feature (i), Data Set (t)) i=i+1 Data set (t)=n No Predict IM(Feature (i)) Yes i=i+1 Split Data Sets 1,,n based on Max Predicted IM (Feature(1), Feature (k)) For Each Subset, P (Class 1) Yes End TREE, Classify Irrelevant Features Manohar, Tucker

28 Split Data Sets (t 1,,t n ) : Max IM Time t 1 Time t n A i,1 A i,k Split Data Sets (1,..,n) based on k mutually exclusive Feature values of Feature A i Research Methodology Manohar, Tucker

29 n- time stamped data sets No IM(Feature (i), Data Set (t)) i=i+1 Data set (t)=n No Predict IM(Feature (i)) Yes i=i+1 Split Data Sets 1,,n based on Max Predicted IM (Feature(1), Feature (k)) For Each Subset, P (Class 1) Yes End TREE, Classify Irrelevant Features Manohar, Tucker

30 n- time stamped data sets No IM(Feature (i), Data Set (t)) i=i+1 Data set (t)=n No Predict IM(Feature (i)) Yes i=i+1 Split Data Sets 1,,n based on Max Predicted IM (Feature(1), Feature (k)) For Each Subset, P (Class 1) Yes End TREE, Classify Irrelevant Features Manohar, Tucker

31 Data Mining Predictive Model Time t 1 Time t n Threat Results Manohar, Tucker

32 Proposed Methodology Research Methodology Manohar, Tucker

33 Step 4: Decision Support Early Warning System (EWS) is a graphical user interface (GUI) that display the percentage probability of threat/violent action being committed. Research Methodology Manohar, Tucker

34 APPLICATION CASE STUDY Case Study Manohar, Tucker

35 Possible Threat Scenario Case Study BBC UK (2008) Manohar, Tucker

36 CASE STUDY: TEST DATA Voluntary participants from the University community were invited to enact the threat and non-threat actions Recreated in an indoor space, similar to a high profile speech The data collected is then used to train the predictive models The study was approved by the IRB and the ORP at the Pennsylvania State University, University Park campus, under the title A Dynamic Pattern Recognition Framework for Mining and Predicting Emerging Threats and is filed as IRB # Study: 24 Subjects spanning 2 months Case Study Manohar, Tucker

37 THREAT PREDICTION RESULTS Low level threat prediction High level threat prediction Results Manohar, Tucker

38 RESULTS Confusion matrix for REPTree: Confusion matrix for Naive Bayes: Accuracy of Ensemble Methods: 86.8% Class FALSE TRUE Class FALSE TRUE FALSE FALSE TRUE TRUE Accuracy measures for REPTree: Accuracy measures for Naïve Bayes: Accuracy Precision Recall F- Measure PRC Area ROC Area Accuracy Precision Recall F- Measure PRC Area ROC Area 95.3% 96.9% 97.7% 97.3% 99.1% 96.9% 82.7% 87.3% 93.1% 90.1% 90.9% 71.8% Results Manohar, Tucker

39 CONCLUSION AND FUTURE WORK Conclusion and Future Work Manohar, Tucker

40 Conclusion and Future Work The most common surveillance systems today are reactive in nature and are not capable of actively predicting the emergence of a threat by analyzing past data collected. Privacy preserving data mining methodology This methodology takes the first step towards addressing these issues while providing promising results Expand the definition of threat Conclusion and Future Work Manohar, Tucker

41 Contributors: Dr. Conrad S. Tucker, D.A.T.A. Lab members, Research Participants from PSU. References: ACKNOWLEDGEMENTS AND REFERENCES 1. Quinlan, J. R. C4.5: Programs for Machine Learning. Morgan Kaufmann Publishers, Joshi, Karuna Pande. "Analysis of data mining algorithms." University of Minnesota. Retrieved July 25 (1997): J. Han, M. Kamber, J. Pei, Data Mining: Concepts and Techniques, Third edition, Data-Driven Decision Tree Classification for Product Portfolio Design Optimization, Conrad S. Tucker and Harrison M. Kim, J. Comput. Inf. Sci. Eng. 9, (2009), DOI: / J. L. Raheja, A. Chaudhary, K. Singal, Tracking of fingertips and centers of palm using KINECT, International Conference on Computational Intelligence, Modeling & Simulation, 2011, Ya-Li Hou and Grantham K.H. Pang, Human detection in crowded scenes, IEEE international conference on image processing, 2010, References Manohar, Tucker