Reflective Database Access Control Lars Olson Ph.D. Thesis Defense.

Reflective Database Access Control

Lars OlsonPh.D. Thesis Defense

Introduction

Database

Alice Bob Carol David

ACM-Based Access Control

Employees

Name SSN Salary Dept Position

Alice 123456789

80000 HR CPA

Bob 234567890

70000 Sales Sales Rep

Carol 345678901

90000 Sales Manager

David 456789012

90000 HR Manager

ACM Entrie

s

Alice

David


Employees

Name SSN Salary Dept Position

Alice 123456789

80000 HR CPA

Bob 234567890

70000 Sales Sales Rep

Carol 345678901

90000 Sales Manager

David 456789012

90000 HR Manager


Sales_Employees

Bob Sales

SalesCarol

Sales Rep

Manager

ACM Entrie

s

Bob

Carol

ACM Weaknesses

• Complicated policies can be awkward to define

• “Every employee can access their own records”

• “Every employee can view the name and position of every other employee in their department”

Motivation

• ACMs describe extent, rather than intent

• Decision support data is often already in the database– Redundancy– Possibility of update anomalies

Reflective Database Access Control

• Solution: access policies should contain queries– Not limited to read-only operations– Policies not assumed to be “omniscient”

• Is this a secure solution? (CCS ’08)• Is this a practical solution? (DBSec ’09)• What is it useful for? (SPIMACS ’09)

Database

Thesis Statement

Datalog-based reflective database access control can provide a flexible, scalable, and efficient mechanism for defining, enforcing, and formally reasoning about fine-grained access control policies.

Outline

• Challenges for RDBAC• Theory

– Formalism using Transaction Datalog– Security analysis

• Implementation– Prototype description– Evaluation

• Case Studies– Medical database– Building automation system

• Future Work and Conclusion

Outline






Application-Layer Security

Database

Application A

Access Control Rules

User a

User b

User c

A

Oracle Virtual Private Database

• User-defined function as query filter– Access to current user– Access to other table data (excluding

current table)– Non-omniscient— subject to policies

protecting other data

• Flexible— a little too flexible…

create or replace function leakInfoFilter (p_schema varchar2, p_obj varchar2)

return varchar2 asbegin

for allowedVal in (select * from alice.employees) loop

insert into logtable values (sysdate,'name:' || allowedVal.name|| ', ssn:' || allowedVal.ssn|| ', salary:' || allowedVal.salary);

end loop;commit;return '';

end;

Pitfalls in Reflective AC

Not Necessarily a Problem

• Note:– Only privileged users can define VPD policies.– Using POLICY_INVOKER instead of SESSION_USER in the employees table would solve this problem.

• Still, centralized policy definers not ideal– Scalability– Difficulty in understanding subtle policy

interactions…and you have to deal with surly DB admins

Pitfalls in Reflective AC

• Queries within policies must be executed under someone’s permissions.

• Cyclic policies cause infinite loop.• Long chains of policies may use the

database inefficiently.• Determining safety is undecidable, in

general.

Desirable Properties

• Policy can depend on user attributes or object attributes in database

• Updates immediately affect policy evaluation• Policies are fine-grained• Policies may modify database• Lower-privileged users may define privileges

for their own tables (non-omniscient policies)• Model has formal mathematical basis• System performance is comparable to current

technology

Outline






Transaction Datalog

• Datalog extended with assertion and retraction semantics

• Inference process extended to track modifications

• Concurrency and atomicity• Implicit rollback on failure

Transaction Datalog Example

• State:emp(alice, 1234, 80000, hr, manager).emp(bob, 2345, 60000, hr, accountant).

• Transaction Base:changeSalary(Name, OldSalary, NewSalary) :- emp(Name, SSN, OldSalary, Dept, Pos), del.emp(Name, SSN, OldSalary, Dept, Pos), ins.emp(Name, SSN, NewSalary, Dept, Pos).

• Runtime queries:changeSalary(alice, 50000, 100000)? No.changeSalary(alice, 80000, 100000)? Yes.

• Allow users to access their own records:view.emp(User, Name, SSN, Salary, Dept, Pos) :- emp(Name, SSN, Salary, Dept, Pos), User=Name.

• Allow users to view names of employees in their own department:view.emp(User, Name, null, null, Dept, Pos) :- emp(User, _, _, Dept, _), emp(Name, _, _, Dept, Pos).

TD as a Policy Language

TD as a Policy Language

• Restrict and audit sensitive accesses:view.emp(User, Name, SSN, Salary, Dept, Pos) :- emp(User, _, _, hr, _), emp(Name, SSN, Salary, Dept, Pos), ins.auditLog(User, Name, cur_time).

• Chinese Wall policy:view.bank1(User, Data1, Data2) :- cwUsers(User, 1, OldValue), bank1(Data1, Data2), del.cwUsers(User, 1, OldValue), ins.cwUsers(User, 1, 0).

Fixing the Leak

• Policies must always run under the definer’s privileges:view.a(User, ...) :- view.b(alice, ...), view.c(alice, ...).

• Basic table owner privileges can be generated automatically.view.a(alice, ...) :- a(...).

Formal Safety Analysis

• Efficiency of answering the question “Can user u ever gain access right r to object o?”– Excludes actions taken by trusted users

• TD can implement HRU model• Consequence: safety is undecidable

in general

Decidable Class #1

• Read-only policies

• Check whether subject s can access object o initially

• Ignore irrelevant tables• Infrequent updates

– Polynomial-time safety check– Unsafe configurations can be rolled back

• Retraction-free• “Safe rewritability”

– Rewrite policies to calculate their effect on the database, e.g.:

• Original policy rule:p(X) :- q(X, Y), ins.r(X, Y), s(Y, Z).

• Rewritten rules:r(X, Y) :- q(X, Y).

p(X) :- q(X, Y), r(X, Y), s(Y, Z).

– Rewritten rules must be range-restricted to ensure efficient computation

Decidable Class #2

Proving Safety Decidability

• Database never shrinks• Rewritten rules provide upper bound on

database• Every sequence of operations reaches

fixed point• Finitely many operations

• Too ugly?– Use upper bound as conservative estimate– No negation semantics in TD

Outline






System Architecture

Database

TD Policy

Individual User-defined Policies

Policy Compiler

SQL:1999 Recursive

View Definition

s

Schem

a

met

adat

a

User queries normally

Compilation to SQL Views

• Off-the-shelf SQL databases benefit from years of query optimization research

• Datalog, SQL roughly equivalent– User ID provided by CURRENT_USER system

variable– Recursion requires SQL:1999

• Assertions and retractions– SQL syntax does not permit insert or delete within select statement

– Execution ordering is significant

Side-Effects Within Queries

• Ideally, part of the language– Transaction control– Variable bindings

• In practice, executed as UDF– Execution ordering depends on query plan

• Executing UDF(s) last• Forbids policies with mid-execution side-

effects

– Requires separate connection setup in DBs that do not support side-effects

Compilation Process (1st Pass)

view.emp(User, Name, SSN, Salary, Dept, Pos) :-view.emp('alice', User, _, _, 'hr', _),view.emp('alice', Name, SSN, Salary, Dept, Pos),view.ins.auditLog('alice', User, Name, cur_time).

with view_emp as (...union all

select e1.Name as User,e2.Name as Name, ..., e2.Pos as Pos,1 as Assert_flag,e1.Name as Assert_param1,e2.Name as Assert_param2

from view_emp e1, view_emp e2where e1.Dept = 'hr' and e1.Name =

'alice' and e2.Name = 'alice'union all...)select distinct User, Name, ..., Posfrom view_empwhere Assert_flag = 0 or (Assert_flag = 1and assert_auditLog(Assert_param1,

Assert_param2) != 0)

function assert_auditLog(@User varchar,@Name varchar)

...

Schema:

User, Name, SSN, Salary, Dept, Pos,

Assert_flag, Assert_param1, Assert_param2

Compilation Process (2nd Pass)

function assert_auditLog(@User varchar,@Name varchar)

...

with view_emp as (...union all

select e1.Name as User,e2.Name as Name, ..., e2.Pos as Pos,1 as Assert_flag,e1.Name as Assert_param1,e2.Name as Assert_param2

from view_emp e1, view_emp e2where e1.Dept = 'hr' and e1.Name =

'alice' and e2.Name = 'alice'union all...)select distinct User, Name, ..., Posfrom view_empwhere Assert_flag = 0 or (Assert_flag = 1and assert_auditLog(Assert_param1,

Assert_param2) != 0)

Schema:

User, Name, SSN, Salary, Dept, Pos,

Assert_flag, Assert_param1, Assert_param2

view.emp(User, Name, SSN, Salary, Dept, Pos) :-view.emp('alice', User, _, _, 'hr', _),view.emp('alice', Name, SSN, Salary, Dept, Pos),view.ins.auditLog('alice', User, Name, cur_time).

Compilation Process (cont.)

• Filter on user:

create view view_emp_public asselect Name, ..., Posfrom view_empwhere User = CURRENT_USER;

grant select on view_emp_public to public;

Optimizations

• Recursive views are expensive!• Use predicate unfoldingview.emp('alice', Name, SSN, Salary, Dept, Pos) :-emp(Name, SSN, Salary, Dept, Pos).

…allows us to rewriteview.emp('alice', User, _, _, 'hr', _)

…toemp(User, _, _, 'hr', _)

Optimizations (cont.)

• union all is expensive (although not as bad as recursion)– Build query dynamically– Pre-compute portions of rule– If rule doesn’t apply, we can eliminate a union

– Simulated with stored procedure

Evaluation

• Baseline– Custom-defined views– ACM-based enforcement– Two baselines for side-effect queries

• No side-effect• Side-effect UDF called within view

• Compiled views– Unoptimized, with recursion– Optimized with predicate unfolding

• Simulated optimization with predicate unfolding and union all elimination

Timing Results (fixed DB size)

0.01

0.1

1

10

100

1000

10000

100000

HR Manager Insurance ChineseWall

Avg

. E

xecu

tio

n T

ime

(sec

)

Baseline 1

Baseline 2

Recursive

Optimized

Target

Timing Results (fixed query)

0.0001

0.001

0.01

0.1

1

10

100

1000

10000

100000

1000 10000 100000

Database Size

Avg

. Exe

cuti

on

Tim

e (s

ec)

Recursive

Optimized

Target

Baseline 1

Outline






Case Study: Medical Database

• HIPAA legislation– Protects privacy of patients– Access to electronic health records must

be restricted “based on the specific roles of the members of their workforce.”

• Idealism meets reality: emergencies are common

• Commonly implemented by Honor System, e.g. sign a form yearly

Example Policies

• Patients may view their own medical data

• Primary care physicians may view their own patients’ data

• Caregivers assigned to consult with a patient may view that patient’s data

• Current employees may access any patient’s record, but an audit record is generated

Formal Security Analysis

• “No untrusted user can ever gain access to a patient’s lab results.”

• Uses upper-bound estimate on append-only policies– Rules with retractions, rules not safely

rewritable omitted– Sample database populated, verified with

Prolog– Omitted rules analyzed manually

• Analysis scalability– Running time A: increased patients & doctors– Running time B: increated patients only

Formal Security Analysis

1

10

100

1000

10000

100000

10,000 100,000 1,000,000 10,000,000

Number of patients

Exe

cuti

on

Tim

e (s

ec)

Running Time A

Running Time B

Case Study: Building Automation

System

BuildingControlNetwork

Building Resources

Legacy BASControllerand DB

LocalNetwork

ClassRegistration

TeachingAssignments

to Internet

Firewall

Example Policies

• Users who are given delegation privileges over a room may add or delete users that may access the room

• Students enrolled in a class may unlock the room where the class occurs during normal class hours– Attendance recorded– Internet access disabled

• Anyone may purchase items from a vending machine, with cost of items deducted from their account

Outline






Future Research Possibilities

• Improvements to TD– Aggregation– Negation– Atomic policies for updates

• Improvements to analysis– Retraction analysis– State-independent analysis– Information flow using delegated

privileges

Future Research Possibilities

• Further DB integration– Automatic checks for safety– Pre-computing optimization– Side-effects and ordering

• Development of Case Studies– Discretionary access to patient records

• “Trusted users” no longer constant• Specifying exceptions

– Firewall rules

Conclusion

• Reflective Database Access Control is a more flexible model than View-Based Access Control.– Easier to model policy intent– Subtle data interactions create new

dangers• Transaction Datalog provides a

reasonable theoretical basis for RDBAC.– Expressive semantics for describing policy

intent– Safety analysis

Conclusion

• Compilation of TD rules to SQL views implements RDBAC with current database technology.

• Performance cost of compiled views is low and can yet be improved.

• RDBAC provides benefits for real-world scenarios.

Reflective Database Access Control Lars Olson Ph.D. Thesis Defense.

Documents