Secure Data Management: Foundations, Systems and Applications My - - PowerPoint PPT Presentation
Erik Jonsson School of Engineering and Computer Science Distinguished Lecture Series 2017-2018 Secure Data Management: Foundations, Systems and Applications My Journey 1985-Present Dr. Bhavani Thuraisingham The University of Texas at Dallas
Erik Jonsson School of Engineering and Computer Science Distinguished Lecture Series 2017-2018
April 27, 2018
Mathematics and Physics at the University of Ceylon. My husband, also of Tamil origin, was finishing his PhD at the University of Cambridge, England.
moved to the US in 1980 for better opportunities.
my son was a baby, so took visiting faculty positions in Socorro, NM and later Minneapolis, MN for 3 years and then joined Control Data Corp. as a Senior Software Developer in Computer Networks and Distributed Systems.
USAF contract to design a high assurance secure database, and Honeywell interviewed me and wanted to hire me; all three had to come together.
Dallas for 32+ years. * https://www.youtube.com/watch?v=xfBie2oVzkA 2
3
management systems (DBMS) in a protected subsystem as objects of special types are only accessible to subjects executing in the DBMS domain.
unique design.
underlying security policy to meet database requirements.
the DBMS and user domains.
security model.
Informix, and Ingres in the early 1990s. * Bhavani M. Thuraisingham: Security checking in relational database management systems augmented with inference engines. Computers &
Security 6(6): 479-492 (1987) (Landmark paper that spawned research on the inference problem) Paul D. Stachour, Bhavani M. Thuraisingham: Design of LDV: A Multilevel Secure Relational Database Management System. IEEE T
Data Eng. 2(2): 190-209 (1990)
4
5
Response Pipeline Update Pipeline Metadata Pipeline * This research had a significant impact on the National Computer
Security Center’s Purple Book in 1991 – Evaluation Criteria for Secure Database Systems
6
User Interface Manager
Policy Manager Security Policies
Query Processor: Policies during query and release
Update Processor: Policies during update
Database Design Tool Policies during database design
MLS Database
MLS/DBMS
unauthorized information
released
the response should be released
design time
Maneuver Control System in a distributed environment connecting systems in MA, VA, and NJ.
* Bhavani M. Thuraisingham, William R. Ford, Marie Collins, J. O'Keeffe: Design and Implementation of a Database Inference Controller. Data Knowl. Eng. 11(3): 271-297 (1993) Bhavani M. Thuraisingham, William R. Ford: Security Constraints in a Multilevel Secure Distributed Database Management System. IEEE Trans. Knowl. Data Eng. 7(2): 274-293 (1995)
was a significant result in database security in 1990 (Proceedings National Computer Security Conference, 1990)
formulation of K-Anonymity (followed by L-Diversity and Differential Privacy)
* Bhavani Thuraisingham, Recursion Theoretic Properties of the Inference Problem, IEEE Computer Security Foundations Franconia, NH, June 1990 (also available as MTP Technical Report, MTP 291) Bhavani Thuraisingham. On the Complexity of the Privacy Problem in Databases, Data Mining: Theory and Practice, Springer, Editors: TY Lin et al, 2005. First work to integrate computability theory with secure data
7
8
Theorem I (i) For each privacy level L, PP[L] is recursively enumerable (ii) For each privacy level L, PP[L] is either recursive or nonsimple (iii) If all privacy functions which model the rules in deductive databases are deterministic, then for each privacy level L, PP[L] is either recursive or a cylinder (iv) If the privacy level L1 dominates the privacy level L2, then PP[L1] PP[L2]where is the subset function Privacy problem: The privacy problem with respect to privacy level L is the set of all quadruples <B, F, C, A> such that there is some x belongs to CnF(B) and the privacy level of x dominates L. Note that we assume that the set of privacy levels form a lattice. Formally stated the privacy problem at level L is the set: PP[L] = {<B, F, C, A> | Level (B) ≤ L and x (x CnF(B) and Level(x) > L)} where is the “there exists” symbol. Multilevel Database: A multilevel deductive database is a quadruple <B, F, C, A> where B is a database, F is a privacy function, C is a recursive set of privacy policies and A is an algorithm (i.e. an effective procedure) which assigns privacy levels to the data based on C. (Note that since C is recursive, one can effectively decide whether a privacy constraint belongs to C.) Theorem 2 (i) There is a situation where PP[Public] is not recursive. (ii) Assuming that the privacy functions are deterministic, there is a situation where PP[Public] is not creative. (iii) There is a situation where PP[Public] is neither recursive nor a cylinder.
9
(u,0) (u,1) (u,2) (u,v) (u,v+1) (u,v+2) T(e,u,v) (0,0) (0,1) (0,2) (0,u+v) (0,u+v+1) (0,u+v+2) (u,0) (u,1) (u,2) (u,v) (u,v+1) (u,v+2) T(e,u,v) (0,0) (0,1) (0,2) (0,u+v) (0,u+v+1) (0,u+v+2)
We first show that given a recursively enumerable set W, there is a situation S such that W m PP[Public]. Note that m is the many-one equivalence relationship. The result is then immediate from the following reasoning. * It has been shown that there is set K which is creative. K is the set {x: the xth partial recursive function halts on input x} * The situation S that is constructed from the recursively enumerable set K will guarantee that PP[Public] is creative. This is because if the two sets A and B are many one equivalent and A is creative, then so is B. Therefore, if PP[Public] is creative then it cannot be recursive. Given a recursively enumerable set W, we create a situation S by defining a set of privacy constraints and a privacy function. Let the set of privacy constraints be {(0,0)}. That is, the only element that is assigned the private level is the pair (0,0). We consider pairs of natural numbers. This does not cause any problem due to the existence of the pairing function from NxN onto N where N is the set of all natural numbers. The set of privacy constraints is recursive (note that in this case it is also finite) and does not depend on W. We define a privacy function, which depends on W as follows. We assume that e is the index of W. The privacy function f for a pair (u,v) is defined as follows: {(u, v+1)} if u 0 AND NOT T(e, u, v) f(u,v) = {(0, u+v) if u 0 AND T(e, u, v) {(u, v-1) if u = 0 AND v 0 (the empty set) if u = 0 AND v = 0 Note that T is the Kleene’s T Predicate
Graphical Representation of the Privacy Function f
PP[Public] = {(u,v): there is a path via f from (u,v) to (0,0)}
We have shown that We m PP[Public] where We is the eth recursively enumerable set. Note that m is many-one equivalence.
multilevel databases at that time.
multilevel databases.
multilevel database systems.
controller and patented and licensed.
* Bhavani M. Thuraisingham: A Nonmonotonic Typed Multilevel Logic for Multilevel Secure Database/Knowledge-Based Management
Bhavani M. Thuraisingham: A Nonmonotonic Typed Multilevel Logic for Multilevel Secure Database/Knowledge-Based Management Systems II IEEE CSFW 1992
10
proper axioms, and a set of inference rules.
an NTML theory whose proper axioms are the elementary facts, the schemas and the general laws.
formula) of the NTML theory.
closed (i.e., with no free variables), and an open query is one with at least one free variable.
at level L. Then evaluating the query Q amounts to proving that (W,L) is a theorem of the NTML theory.
11
12
and our Project Coordinator: Ms. Rhonda Walls
13
Influenced by the 9/11 Commission Report *
14
information has been lost (trustworthy partners)
trustworthy partners
partner
monitoring them so that we can defend ourselves from dynamic situations * Bhavani Thuraisingham, Assured Information Sharing: Technologies, Challenges and Directions. Intelligence and
Security Informatics 2008: 1-15
15
Policies Ontologies Rules In RDF
JENA RDF Engine RDF Documents
Policy/Inference Engine/ Rules Processor e.g., Pellet
Interface to the Semantic Web
Technology by UT Dallas
Barbara Carminati, Elena Ferrari, Raymond Heatherly, Murat Kantarcioglu, Bhavani M. Thuraisingham: A semantic web based framework for social network access control. SACMAT 2009: 177-186 Bhavani M. Thuraisingham, Tyrone Cadenhead, Murat Kantarcioglu, Vaibhav Khadilkar: Secure Data Provenance and Inference Control with Semantic
Research challenges we addressed included
evaluation
User Interface Layer RDF Data Preprocessor Policy Translation and Transformation Layer MapReduce Framework for Query Processing Hadoop HDFS Agency 1 Agency 2 Agency n
…
RDF Data and Policies SPARQL Query Result
16
Example Policies Agency 1 wishes to share its resources if Agency 2 also shares its resources with it Agency 1 shares a resource with Agency 2 provided Agency 2 does not share with Agency 3
*Tyrone Cadenhead, Murat Kantarcioglu, Vaibhav Khadilkar, Bhavani M. Thuraisingham: Design and Implementation
Mohammad Farhan Husain, James P. McGlothlin, Mohammad M. Masud, Latifur R. Khan, Bhavani M. Thuraisingham: Heuristics- Based Query Processing for Large RDF Graphs Using Cloud Computing. IEEE Trans. Knowl. Data Eng. 23(9): 1312-1327 (2011)
prove that the system meets the specifications.
systems.
metamodel which we then enrich with a rewrite-based operational semantics from which prototype implementations can be directly derived.
17
*M. Fernandez, M. Kantarcioglu, B. Thuraisingham, A
Framework for Secure Data Collection and Management for Internet of Things Proceedings IEEE ACSAC Conference Workshop on Industrial Control Systems, Los Angeles, CA, December 2016.
Proceedings ACM CODASPY/ABAC Symposium, Tempe, AZ, March 2018.
Fernandez, Privacy-Aware Quantified Self Data Management Framework, Proceedings ACM SACMAT, Indianapolis, June 2018. Future Funding possibility through NSA SoS Lablet
Our main focus has been on Category-based Metamodel for Secure Data Collection (CBDC). Future work will include secure data storage, management and sharing. Category: class, group or domain to which entities belong
18
Entities:
e.g. individual sensors, aggregators, clocks, etc.
and also contextual information (such as location, time, identifier, etc.)
DICAd ⊆ DId × C, such that (di , c) ∈ DICAd iff the data item di ∈ DId , generated by the device d ∈ Dev , is assigned to the category c.
(a, c) ∈ ACA iff action a ∈ A can be performed on data items assigned to the category c.
iff action a ∈A is banned for data items assigned to the category c ∈ C.
iff action a ∈ A is authorized on the data item di generated by d ∈ Dev.
iff action a ∈A is banned on data item di generated by d ∈ Dev .
that the action a is neither authorized nor banned on the data item di emanating from d .
19
A category-based data collection policy is a tuple (E , {DICAd }d∈ Dev , ACA, BACA, {ADId }d∈ Dev , {BADId }d∈ Dev , {UN DETd }d∈ Dev ) where E = (Dev , DI, A, C, S, ⊆ ), such that axioms (dc1)-(dc4) are satisfied. Operational semantics: axioms (dc1)-(dc4) can be realized through a set of functions, defined by rewrite rules. (dc1/) ADId (A, Di ) → if A ∈ ACA∗(below(DICAd (Di ))) then accept else if A ∈ BACA∗(above(DICAd (Di ))) then forbid else undetermined 20
Rental prices vary depending on whether the truck is taken out of the country or not. Drivers who are not planning to leave the country can benefit from a discount. Trucks fitted with tracking devices able to transmit GPS locations. Goal: tracking information transmitted only if truck crosses border.
Next Steps: Develop an Implementation and Check for Efficiency; Develop a CBAC Model for Data Sharing
21
first at IFIP 11.3 in 1996 and later at PAKDD in 1998 and subsequently published a landmark paper* while at NSF in 2002 that spawned a new area of research.**
Trees and the Perturbation Method*** between 2005 - 2008.
hosted an NSF Workshop on Big Data Security and Privacy in September 2014 and presented the results to the Interagency Working Group in February 2015.
* Bhavani M. Thuraisingham: Data Mining, National Security, Privacy and Civil Liberties. SIGKDD Explorations 4(2): 1-5 (2002) ** https://www.utdallas.edu/~bxt043000/Press-Releases/Bhavani-MITRE-article-with-Marty-Faga.pdf **Li Liu, Murat Kantarcioglu, Bhavani M. Thuraisingham: The applicability of the perturbation based privacy preserving data mining for real-world data. Data Knowl. Eng. 65(1): 5-21 (2008)
22
Noise Addition X (original) X’ ( noise added) Modified Data Mining Process Final Modified Result
Network traffic Classification model Attack traffic Firewall Block and quarantine Benign traffic Server Expert analysis and labeling
23
Big Data Streams:
*Mehedy Masud, Latifur Khan and Bhavani Thuraisingham, Data Mining Tools for
Malware Detection, CRC Press, 2011. Mohammad M. Masud, Jing Gao, Latifur Khan, Jiawei Han, Bhavani M. Thuraisingham: Classification and Novel Class Detection in Concept-Drifting Data Streams under Time
evolving behaviors tend to mask such anomalies.
supervised learning that addresses this challenge by maintaining an evolving collection of multiple models to classify dynamic data streams of unbounded length.
insider threat streams relative to traditional supervised learning (traditional SVM and one-class SVM) and other single-model approaches.
that maintains a compressed dictionary of repetitive sequences found throughout dynamic data streams of unbounded length to identify anomalies.
sequences.
streams containing insider threat anomalies.
learning approaches and boosts the effectiveness over supervised learning approaches. * Sponsor: AFOSR 24
Feature Extraction & Selection
Anomaly?
j
System Call System Call
chunki+1 chunki
System log
Testing on Data from Chunki+1
Online learning
Gather Data from Chunki Feature Extraction & Selection Learning Algorithm
Supervised - One class SVM, OCSVM Unsupervised - Graph based Anomaly Detection, GBAD
Ensemble-based Stream Analytics Ensemble of Models Update models
* Pallabi Parveen, Nate McDaniel, Jonathan Evans, Bhavani Thuraisingham, Kevin W. Hamlen and Latifur Khan, "Evolving Insider Threat Detection Stream Mining Perspective", "Journal International Journal on Artificial Intelligence Tools," World Scientific Publishing, 2013
25
system calls, etc.
indicating an insider threat
We need to collect these repeated sequences of commands in an unsupervised fashion
Overcome: Generating a LZW dictionary with combinations
possible potential patterns in the gathered data using Lempel-Ziv-Welch algorithm (LZW)
Overcome: Compress the dictionary
Pallabi Parveen, Nate McDaniel, Varun S. Hariharan, Bhavani M. Thuraisingham: Unsupervised Ensemble Based Learning for Insider Threat
26
𝑗𝑘 is the frequency of the pattern i in chunk j.
𝑗𝑘
27
28
It is not concept drift It is not online learning Adversary adapts to avoid being detected During training time (i.e., data poisoning) During test time (i.e., modifying features when data mining is deployed) There is game between the data miner and the adversary Understanding Adversarial Learning
29
*Y. Zhou, M. Kantarcioglu, B. Thuraisingham, and B. Xi, Adversarial support vector
machine learning, ACM SIGKDD ‘12 One of the early efforts on Adversarial Machine Learning for Cyber Security
30
Free-range attack
Adversary can move malicious data anywhere in the feature space The only knowledge the adversary needs is the valid range of the feature An attack is bounded in the following form.
For every j in [1, d] 𝐷
𝑔 is between 0 and 1 and controls the aggressiveness of the attacks; 0
means no attack and 1 means the most aggressive attack 𝑦𝑘 (max/min) is the largest/smallest value of the jth feature of a data point 𝑦𝑗 (which is 𝑦𝑗𝑘) can take.
Other models: Restrained attack
Assume that the Adversary would be reluctant to let a data point move fare
away from its original position
Reason is that great displacement often entails loss of malicious utility
Developed an Adversarial SVM for each attack
min - xij) £dij £ C f (x. j max - xij)
31
black dashed line is the standard SVM classification boundary, and the blue line is the Adversarial SVM (ADV-SVM) classification boundary 6
Threat Model Example
Test time/ Deployment Time Attacks
Attacker modifies x to x’ Modify packet length by adding dummy bytes Add good word to spam e-mail Add noise to an image
32
space
would not cope well with attacks that modify data to change the makeup of the sample space
considered.
understanding of the attacker.
33
flexible security policies and addresses malware attacks (e.g., SQL injection, covert channels).
Tolerance and Real-time Processing
34
Data Management and the Inference Problem
Ingres, Ontos)
System
2005-2015
Interagency working group) resulting in multiple research programs, 2015
2017- present
35
Walls, and all our students as well as members of the Cyber Security Research and Education Institute at UT Dallas
36
* Part of my strong support group of women over the 32+ years
Academic Excellence in Cyber Security Research, June 2008 (CAE-R); Recertifications in 2014 until 2021.
Usenix Security, ACM KDD, PVLDB, IEEE ICDE, IEEE ICDM, IEEE Big Data, - - - - )
(Gel, Lary et al)- multiple joint funded programs from NSF and DoD as well as papers.
37
38
transfer, operational systems, publications, patents, books, keynote addresses, leadership roles, and more importantly an H Index of 52 and Citation Count of 10,366 as of April 26, 2018.
discouraged; Only your enemies will benefit from your downfall.
me for 43 years, (ii) My mother who always said “Clever Girl” regardless of whether I came first
my father at 16.
39
WiCyS and WiDS
premier organization and integrate Cyber Security with Data Science
continue to focus also on education proposals including NRT and ADVANCE
where one can truly build a community – like I did in Data Mining, Security and Privacy at NSF 2001-2004.
Data Security Integrated with Software Development
40