- Students records (personal data, registration to examinations, grades)
- Services:
- Enrolment/expulsion of students
- Registration to examination
- Registration of examination marks
- Information and attestations desk
- Operational Risks
- Conditio sine qua non: Provability of information properties
- Fake registration to examinations: integrity, non-repudiability ("nicht-abstreitbar")
- Leakage of grades, personal data: confidentiality, integrity
- Forgery of attestations: authenticity, integrity
Industry Control Systems
- e.g. Factorys, energy and water plants (public infrastructure)
- "Chinese Hacking Team Caught Takin over decoy water plant"
- "Internet Attack shuts off the Heat in Finland"
- Operational risks: Integrity & Availability of public community support systems
[Self Study Task]() Read about these two scenarios. Find one or more recent examples for attacks on public infrastructure, including some technical details, in the news. Keep all these scenarios in mind, we will come back to them in the next chapter:
- [Hacker breached 63 universities and government agencies](https://www.computerworld.com/article/3170724/hacker-breached-63-universities-and-government-agencies.html)
- [Ransomeware attacks on public services](https://www.nytimes.com/2019/08/22/us/ransomware-attacks-hacking.html)
- [Worst data leaks and breaches in the last decade](https://www.cnet.com/how-to/14-of-the-worst-data-leaks-breaches-scrapes-and-security-snafus-in-the-last-decade/)
### Message
- Goal of IT Security: **Reduction of Operational Risks of IT Systems**
- Security requirements, which definewhatsecurity properties a system should have.
- These again are the basis of asecurity policy: Defineshowthese properties are achieved
Influencing Factors
- Codes and acts (depending on applicable law)
- EU General Data Protection Regulation (GDPR)
- US Sarbanes-Oxley Act (SarbOx)
- Contracts with customers
- Certification
- For information security management systems (ISO 27001)
- Subject to German Digital Signature Act (Signaturgesetz), toCommon
- Criteria
- Company-specific guidelines and regulations
- Access to critical data
- Permission assignment
- Company-specific infrastructure and technical requirements
- System architecture
- Application systems (such as OSs, Database Information Systems)
General Methodology: How to Come up with Security Requirements
Specialized steps in regular software requirements engineering:
1. Identify and classifyvulnerabilities.
2. Identify and classifythreats.
3. Match both, where relevant, to yieldrisks.
4. Analyze and decide which risks should bedealt with.
-> Fine-grained Security Requirements
## Vulnerability Analysis
Goal: Identification of
- technical
- organizational
- human
vulnerabilities of IT systems.
> Vulnerability
>
> Feature of hardware and software constituting, an organization running, or a human operating an IT system, which is a necessary precondition for any attack in that system, with the goal to compromise one of its security properties. Set of all vulnerabilities = a system’sattack surface.
### Human Vulnerabilities
Examples:
- Laziness
- Passwords on Post-It
- Fast-clicking exercise: Windows UAC pop-up boxes
- Social Engineering
- Pressure from your boss
- A favor for your friend
- Blackmailing: The poisoned daughter, ...
- An important-seeming email
- Lack of knowledge
- Importing and executing malware
- Indirect, hidden information flowin access control systems
> Social Engineering
>
> Influencing people into acting against their own interest or the interest of an organisation is often a simpler solution than resorting to malware or hacking.
> Both law enforcement and the financial industry indicate that social engineering continues to enable attackers who lack the technical skills, motivation to use them or the resources to purchase or hire them. Additionally, targeted social engineering allows those technically gifted to orchestrate blended attacks bypassing both human and hardware or software lines of defence. [Europol](https://www.europol.europa.eu/crime-areas-and-trends/crime-areas/cybercrime/social-engineering)
Real Cases
> Self Study Task
>
> Investigate the following real-world media (all linked from moodle). Find any potential security vulnerabilities there and give advice how to avoid them.
> - Watch (no listening required) the interview with staff of the French TV station TV5Monde
> - Read the lifehacker article about Windows UAC
> - Read the Washington Times article about Facebook Scam
> - Read the two eMails I received lately.
#### Indirect Information Flow in Access Control Systems
A More Detailed Scenario
- AlphaCompany has two departments: Research & Development(R&D) and Sales
- Ann is project manager and Bob is developer working in R&D on ProjectX, Chris is a busybody sales manager writing a marketing flyer about ProjectX
- All R&D developers communicate via an electronic bulletin board, including any preliminary product features not yet ready for release
- Bob is responsible for informing sales about release-ready features, using ashared web document
> Security Requirement
>
> No internal information about a project, which is not approved by the project manager, should ever go into the product flyer.
Access Control Configuration
- 3 users:ann,bob,chris
- 2 groups:
- crewx: ann, bob, ...
- sales: ann, bob
- Settings:
```
drw- --- --- 1 ann crewx 2020-04-14 15:10 ProjectXFiles
-rw- r-- --- 1 ann crewx 2020-04-14 15:10 ProjectXBoard
-rw- r-- --- 1 bob sales 2020-04-14 14:22 NotesToSales
-rw- --- --- 1 chris sales 2020-04-13 23:58 SalesFlyer.pdf
```
- Result:
- all users apparently set their permissions perfectly – from their own point of view
- all three together createda severe information flow vulnerability...
- Ann has read access to the folder ProjectX Files
- Ann legitimately writes news from these files to the ProjectX Board
- Bob legitimately updates NotesToSales with these news
- Human vulnerability: Bob’s laziness, friendship with Chris, blackmailing by Chris, ... (see above) make him write about unapproved new features
- -> Chris misuses this information in the Sales Flyer...
> Forbidden Information Flow
>
> Internal information about ProjectX goes into the product flyer!
Problem Analysis:
- Limited knowledge ofusers
- limited horizon: knowledge about the rest of a system configuration for making a sound decision about permissions
- limited problem awareness: see “lack of knowledge”
- limited skills
- Problem complexity -> effects ofindividualpermission assignments by users
- Re-designed on a daily basis(many security mechanisms of today’s systems are older than 40 years)
- -> contain conceptual weaknesses and vulnerabilities
#### Buffer Overflow Attacks
Example for Exploitation of Implementation Errors
in privileged system software:
- Operating Systems (OSs)
- SSH demons
- Web servers
- Database servers
Consequence: Privileged software can be tricked into executing attacker’s code
Approach: Cleverly forged parameters overwrite procedure activation frames in
memory
- -> exploitation of missing length checks on input buffers
- -> buffer overflow
What an Attacker Needs to Know
##### Necessary Knowledge and Skills
- Source code of the target program (e. g. a privileged server), obtained by disassembling
- Better: symbol table, as with an executable not stripped from debugging information
- Even better: most precise knowledge about the compiler used w.r.t. runtime management
- how call conventions affect the stack layout
- degree to which stack layout is deterministic, which eases experimentation
Sketch of the Attack Approach (Observations during program execution)
- Stack grows towards the small addresses
- -> small whenever a procedure is called, all its information is stored in aprocedure frame = subsequent addresses below those of previously stored procedure frames
- in each procedure frame: address of the next instruction to call after the current procedure returns (ReturnIP)
- after storing the ReturnIP, compilers reserve stack space for local variables -> these occupy lower addresses
##### Preparing the Attack
Attacker carefully prepares an input argument msg:`\0 ...\0 /bin/shell#system `
```cpp
void processSomeMsg(char *msg, int msgSize){
char localBuffer[1024];
int i=0;
while (i<msgSize){
localBuffer[i] = msg[i];
i++;
}
...
}
```
##### Result:
- Attacker made victim program overwrite runtime-critical parts of its stack:
- by counting up to the length of msg
- at the same time writing back over previously save runtime information -> ReturnIP
- After finishing processSomeMsg: victim program executes code at address of ReturnIP =address of a forged call to execute arbitrary programs!
- Additional parameter to this call: file system location of a shell
> Security Breach
>
> The attacker can remotely communicate, upload, download, and execute anything– with cooperation of the OS, since all of this runs with the original privileges of the victim program!
##### Self-Study Task
Do an internet research: Find a (more or less) recent example of a successful buffer overflow attack. Describe as precise as possible what happened during the attack and which programming error made it possible!
Commonplace search engines for news articles, but also web databases of software vulnerabilities such as https://cve.mitre.org/ may help you.
- Human: Social engineering, laziness, lack of knowledge
- Organizational: Rights management, key management, room access
- Technical: Weak protection paradigms, specification and implementation errors
#### Examples
Scenario 1: Insider Attack
- Social Engineering, plus
- Exploitation of conceptual vulnerabilities (DAC),plus
- Professionally tailored malware
Scenario 2: Malware(a family heirloom ...)
- Trojan horses: Executable code with hidden functionality.
- Viruses: Code for self-modification and self-duplication, often coupled with damaging the host.
- Logical bombs: Code that is activated by some event recognizable from the host (e. g. time, date, temperature, pressure, geographic location, ...).
- Backdoors: Code that is activated through undocumented interfaces (mostly remote).
- Ransomware: Code for encrypting possibly all user data found on the host, used for blackmailing the victims (to pay for decryption).
- Worms and worm segments: Autonomous, self-duplicating programs. Originally designed for good: to make use of free computing power in local networks.
Scenario 3: Outsider Attack
- Attack Method: Buffer Overflow
- Exploitation of implementation errors
Scenario 4: High-end Malware:Root Kits
- Goal: Invisible, total, sustainable takeover of a complete IT system
- Method: Comprehensive tool kit for fully automated attacks
1. automatic analysis of technical vulnerabilities
2. automated attack execution
3. automated installation of backdoors
4. automated installation and activation of stealth mechanisms
- Target: Attacks on all levels of the software stack:
- firmware
- bootloader
- operating system (e. g. drivers, file system, network interface)
- system applications (e. g. file and process managers)
- user applications (e. g. web servers, email, office)
- tailored to specific software and software versions found there!
> Self-Study Task
>
> Read about the following malware examples, both historical and up to date. Identify each as a virus, logical bomb, backdoor, ransomware, or worm; think about which necessary vulnerabilit(y|ies) make all of these threats so dangerous.
> - One of the most sophisticated pieces of malware ever discovered:Stuxnet
> - One of the first large-scale malware threats in history:Michelangelo
> - Two cooperating pieces of malware that put at risk numerous public institutional IT systems:Emotet and Ryuk.
#### Root Kits
Step 1: Vulnerability Analysis
- Tools look for vulnerabilities in
- Active privileged services and demons (from inside a network:nmap, from outside: by port scans) -> Discovers:web server, remote access server (sshd), file server (ftpd), time server (ntpd), print server (cupsd),bluetoothd,smbd, ...
- Configuration files -> Discovers: weak passwords, open ports
- Operating systems -> Discovers: kernel and system tool versions with known implementation errors
- Using built-in knowledge base: an automatable vulnerability database
- Result: System-specific collection of vulnerabilities -> choice of attack method andtools to execute
Step 2: Attack Execution
- Fabrication oftailored softwareto exploit vulnerabilities in
- Server processes or system tool processes (demons)
- OS kernel itself
to execute code of attacker withroot privileges
- This code
- First installs smoke-bombs for obscuring attack
- Then replaces original system software by pre-fabricated modules
- servers and demons
- utilities and libraries
- OS modules
- containing
- backdoors (-> step 3)
- smoke bombs for future attacks (-> step 4)
- Results:
- Backdoors allow forhigh-privilege access within fractions of seconds
- System modified with attacker’s servers, demons, utilities, OS modules
- Obfuscation of modifications and future access
Step 3: Attack Sustainability
- Backdoors for any further control & command in
- Servers (e. g.sshdemon)
- Utilities (e. g.login)
- Libraries (e. g.PAM, pluggable authentication modules)
- OS (system calls used by programs likesudo)
- Modificationsof utilities and OS to prevent
- Killing root kit processes and connections (kill,signal)
- Removal of root kit files (rm,unlink)
- Results: Unnoticed access for attacker
- Anytime
- Highly privileged
- Extremely fast
- Virtually unpreventable
Step 4: Stealth Mechanisms (Smoke Bombs)
- Clean logfiles (entries for root kit processes, network connections), e.g. syslog,kern.log,user.log,daemon.log,auth.log, ...
- Modify system admin utilities
- Process management(hide running root kit processes), e.g. ps,top,ksysguard,taskman
- File system (hide root kit files), e.g. ls,explorer,finder
- Network (hide active root kit connections), e.g. netstat,ifconfig,ipconfig,iwconfig
- Substitute OS kernel modules and drivers (hide root kit processes, files, network connections), e.g. /proc/...,stat,fstat,pstat
- Result:Processes, files and communication of root kit become invisible
Risk and Damage Potential:
- Likeliness of success: extremely highin today’s commodity OSs
- Attack methods und techniques: exploiting vulnerabilities
- human
- organizational
- technical
- -> A zoo of threats, practical assistance:
- National (Germany): BSI IT-Grundschutz standards and catalogues
- International:Common Criteria
Attacks on Public Infrastructure revisited:
> Self-Study Task
>
> Take a close look at those example scenarios for attacks on public infrastructure you read and researched about in chapter 1. For all of them, try to answer the following questions:
> - Who was the presumed attacker mentioned in the article? Classify them according to what you learned about attacker types.
> - What was the attack objective? Again, classify based on what you learned in this chapter.
> - How was the attack made possible? Identify the types of vulnerabilities exploited.
- Cloud computing:“Loss of VM integrity” -> contract penalties, loss of confidence/reputation
- Industrial plant control:“Tampering with frequency converters” -> damage or destruction of facility
- Critical public infrastructure:“Loss of availability due to DoS attacks” -> interrupted services, possible impact on public safety (cf. Finnish heating plant)
- Traffic management:“Loss of GPS data integrity” -> maximum credible accident w. r. t. safety
#### General Fact: Damage potential is highly scenario-specific
Example: “Confidentiality breach of database contents”
- Articles in online newspapers
- -> small to mediumdamage due to lost paywall revenues
- Account data of banks
- -> mission-criticalloss of trust
- Plant control data of industrial production facility
- -> mission-criticalloss of market leadership
Depends on diverse, mostly non-technical side conditions
- -> advisory board needed for assessment:engineers, managers, users, ...
#### Occurrence Probability Assessment
Examples for risks:
- Cloud computing:“Loss of VM integrity”
- -> depending on client data sensitivity
- Industrial plant control:“Tampering with frequency converters”
- -> depending on plant sensitivity(cf.Stuxnet: nuclear centifuges)
- Critical public infrastructure:“Loss of availability due to DoS attacks”
- -> depending on terroristic threat level
- Traffic management:“Loss of GPS data integrity”
- -> depending on terroristic threat level
General Fact: Occurrence probability ishighly scenario-specific
Example: “Confidentiality breach of database contents”
- Articles in online newspapers
- -> smallfor articles that are publicly available anyway
- Account data of banks
- -> medium, due to high attack costs compared to potential gain
- Plant control data of industrial production facility
- -> high, due to high financial or political gain
Depends on diverse, mostly non-technical side conditions
- -> advisory board needed for assessment:engineers, managers, users, ...
- Security Server: Manages and evaluates these modules
### Implementation Alternative B
Application-embedded Policy: The security policy is only known and enforced by oneuser program $\rightarrow$ implemented in a user-space application
Application-level Security Architecture: The security policy is known and enforced by several collaborating user programs in anapplication systems $\rightarrow$ implemented in a local, user-space security architecture
Policy Server Embedded in Middleware: The security policy is communicated and enforced by several collaborating user programs in adistributed application systems $\rightarrow$ implemented in a distributed, user-space security architecture
> Please read each of the following scenarios. Then select the statement you intuitively think is most likely:
> 1. Linda is a young student with a vision. She fights against environmental pollution and volunteers in an NGO for climate protection. After finishing her studies ...
> - ... she becomes an attorney for tax law.
> - ... she becomes an attorney for tax law, but in her spare time consults environmental activists and NGOs.
> 2. Suppose during the 2022 football world cup, Germany reaches the knockout phase after easily winning any game in the previous group phase.
> - Germany wins the semi-final.
> - Germany wins the world cup.
> Think twice about your choices. Can you imagine what other people chose, and why?
Goal of Formal Security Models
- Complete, unambiguous representation of security policies for
1. analyzing and explaining its behavior:
- $\rightarrow$ “This security policy will never allow that ...”
- $\rightarrow$ “This security policy authorizes/denies an access under conditions ... because ...”
2. enabling its correct implementation:
- $\rightarrow$ “This rule is enforced by a C++ method ...”
How We Use Formal Models: Model-based Methodology
- Abstraction from (usually too complex) reality $\rightarrow$ get rid of insignificant details e. g.: allows statements about computability and computation complexity
- Precisionin describing what is significant $\rightarrow$ Model analysis and implementation
> Security Model
>
> A security model is a precise, generally formal representation of a security policy.
Model Spectrum
- Models for access control policies:
- identity-based access control (IBAC)
- role-based access control (RBAC)
- attribute-based access control (ABAC)
- Models for information flow policies
- $\rightarrow$ multilevel security(MLS)
- Models for non-interference/domain isolation policies
- $\rightarrow$ non-interference(NI)
- In Practice: Most oftenhybrid models
### Access Control Models
Formal representations of permissions to execute operations on objects, e. g.:
- Reading files
- Issuing payments
- Controlling industrial centrifuges
Security policies describeaccess rules $\rightarrow$ security models formalize them
Taxonomy
> Identity-based access control models (IBAC)
>
> Rules based on the identity of individual subjects (users, apps, processes, ...) or objects (files, directories, database tables, ...) $\rightarrow$ “Ann may read ProjectX Files.”
> Role-based access control models (RBAC)
>
> Rules based on roles of subjects in an organization $\rightarrow$ “Ward physicians may modify electronic patient records (EPRs) in their ward.”
> Attribute-based access control models (ABAC)
>
> Rules based on attributes of subjects and objects $\rightarrow$ “PEGI 18 rated movies may only be streamed to users aged 18 and over.”
> Discretionary Access Control (DAC)
>
> Individual users specify access rules to objects within their area of responsibility (“at their discretion”).
Example: Access control in many OS (e. g. Unix(oids), Windows)
Consequence: Individual users
- enjoy freedom w. r. t. granting access permissions as individually needed
- need to collectively enforce their organization’s security policy:
- competency problem
- responsibility problem
- malware problem
> Mandatory Access Control (MAC)
>
> System designers and administrators specify system-wide rules, that apply for all users and cannot be sidestepped.
Examples:
- Organizational: airport security check
- Technical: medical information systems, policy-controlled operating systems(e. g. SELinux)
Consequence:
- Limited individual freedom
- Enforced by central instance:
- clearly identified
- competent (security experts)
- responsible (organizationally & legally)
##### DAC vs. MAC
In Real-world Scenarios: Mostly hybrid models enforced by both discretionary and mandatory components, e. g.:
- DAC: locally within a project, team members individually define permissions w. r. t. documents (implemented in project management software and workstation OSs) inside this closed scope;
- MAC:globally for the organization, such that e. g. only documents approved for release by organizational policy rules (implemented in servers and their communication middleware) may be accessed from outside a project’s scope.
#### Identity-based Access Control Models (IBAC)
Goal: To precisely specify the rights ofindividual, acting entities.
Systems Security(Peter Amthor, ST 2021) 3–35 3.2.1.1 IBAC | ACF, ACM
> Self-Study Task
>
> Have a look at your (or any) operating system’s API documentation. Find a few examples for
> - operations executable by user processes (e. g. Linuxsyscalls),
> - their arguments,
> - the operating systems resources they work with.
> Try to distinguish between subjects, objects and operations as defined in the classical ACF model. Can you see any ambiguities or inconsistencies w. r. t. the model?
> If you have never worked with an OS API, stick to the simple things (such as reading/writing a file). For Linux, you may find help insection 2 of the online manpages.a ahttp://man7.org/linux/man-pages/dir_section_2.html
##### Access Control Matrix
Access Control Functions in Practice
Lampson [1974] already addresses the questions how to ...
- store in a well-structured way,
- efficiently evaluate, and
- completely analyze an ACF:
> Access Control Matrix (ACM)
>
> An ACM is a matrix $m:S\times O \rightarrow 2^{OP}$, such that $\forall s\in S,\forall o\in O:op\in m(s,o)\Leftrightarrow f(s,o,op)$.
An ACM is a rewriting of the definition of an ACF: nothing is added, nothing is left out (“$\Leftrightarrow$”). Despite a purely theoretical model: paved the way for practically implementing AC meta-informationas
- Middleware platforms(CORBA, Jini/Apache River, Web Services)
- Distributed security architectures (Kerberos)
- whose security mechanisms use one of two implementations:
Access Control Lists (ACLs)
- Columns of the ACM: `char*o3[N] = { ”-”, ”-”, ”rw”, ...};`
- Found in I-Nodes of Unix(oids), Windows, Mac OS
Capability Lists
- Rows of the ACM: `char* s1[K] = { ”-”, ”r”, ”-”, ...};`
- Found in distributed OSs, middleware, Kerberos
What we Actually Model:
> Protection State
>
> A fixed-time snapshot of all active entities, passive entities, and any meta-information used for making access decisions is called theprotection state of an access control system.
> Goal of ACFs/ACMs
>
> To precisely specify a protection state of an AC system.
- $Q$ is a finite set of states (state space), e. g. $Q=\{q_0 ,q_1 ,q_2\}$
- $\sum$ is a finite set of input words (input alphabet), e. g. $\sum=\{a,b\}$
- $\Omega$ is a finite set of output words (output alphabet), e. g. $\Omega=\{yes,no\}$
- $\delta:Q\times\sum\rightarrow Q$ is the state transition function
- $\lambda:Q\times\sum\rightarrow\Omega$ is the output function
- $q_0\in Q$ is the initial state
- $\delta(q,\sigma)=q′$ and $\lambda(q,\sigma)=\omega$ can be expressed through thestate diagram: a directed graph $⟨Q,E⟩$, where each edge $e\in E$ is represented by a state transition’s predecessor node $q$, its successor node $q′$, and a string “$\sigma|\omega$” of its input and output, respectively.
Example: Return “yes” for any input in an unbroken sequence of “a” or “b”, “no” otherwise.
> Self-Study Task
>
> I bought a new bluetooth headset for giving this lecture. It has just one button and an LED, and this is what the manual says:
> - Turn the device on by holding the button for 1 sec.
> - Turn the device off at any time by holding for 5 sec.
> - Turn the device on in bluetooth pairing mode by holding for 5 sec.
> - Switch to pairing mode when on, but not in a phone call, by holding for 1 sec.
> - Leave pairing mode by clicking the button.
> - Click to start a call. While in a call, click to hang up.
> - While on, the LED indicates modes: red during a call, yellow during pairing, green otherwise.
> Define a mealy automaton for using this headset by defining $Q,\sum,\Omega$,and $q_0$ and draw the state diagram for $\sigma$ and $\lambda$.
Using the STS, $\sigma:Q\times\sum\rightarrow Q$ is defined by a set of specifications in the normalized form
$\sigma(q,⟨op,(x_1,...,x_k)⟩)$=if $r_1\in m_q(x_{s1},x_{o1}) \wedge ... \wedge r_m\in m_q(x_{sm},x_{om})$ then $p_1\circle ...\circle p_n$ where
- $q=\{S_q,O_q,m_q\}\in Q,op\in OP$
- $r_1 ...r_m\in R$
- $x_{s1},...,x_{sm}\in S_q$ and $x_{o1},...,x_{om}\in O_q$ where $s_i$ and $o_i$, $1\leq i\leq m$, are vector indices of the input arguments: $1\leq s_i,o_i\leqk$
- $p_1,...,p_n$ are HRU primitives
- Note: $\circle$ is the (transitive) function composition operator: $(f\circle g)(x)=g(f(x))$
Whenever $q$ is obvious or irrelevant, we use a programming-style notation
Interpretation: The structure of STS definitions is fixed in HRU:
- “if”: A conjunction of condition clauses (or just conditions) with the sole semantics “is some right in some matrix cell”.
- “then”: A concatenation (sequential execution) of HRU primitives.
Conditions:
Expressions that need to evaluate “true” for state q as a necessary precondition for command $op$ to be executable (= can be successfully called).
Primitives:
Short, formal macros that describe differences between $q$ and $a$ successor state $q′=\sigma(q,⟨op,(x_1 ,...,x_k)⟩)$ that result from a complete execution of op:
- enter r into $m(x_s,x_o)$
- delete r from $m(x_s,x_o)$
- create subject $x_s$
- create object $x_o$
- destroy subject $x_s$
- destroy object $x_o$
- $\rightarrow$ Each of these with the intuitive semantics for manipulating $S_q, O_q$ or $m_q$.
Note the atomic semantics: the HRU model assumes that each command successfully called is always completely executed!
How to Design an HRU Security Model:
1. Model Sets: Subjects, objects, operations, rights $\rightarrow$ define the basic sets $S,O,OP,R$
2. STS: Semantics of operations (e. g. the future API of the system to model) that modify the protection state $\rightarrow$ define $\sigma$ using the normalized form/programming syntax of the STS
3. Initialization: Define a well-known initial stateq $0 =⟨S_0 ,O_0 ,m_0 ⟩$ of the system to model
- Rights: Exactly one right allows to execute each operation: $R\congruent OP$
- $\rightarrow R=\{write, read\}$
2. State Transition Scheme
- Effects of operations on protection state:
- writeSolution
Informal Policy: “A sample solution (...) can be downloaded by students only after submitting their own solution.” $\Leftrightarrow$ “If the automaton receives an input ⟨writeSolution,(s,o)⟩ and the conditions are satisfied, it transitions to a state where s is allowed to download the sample solution.”
```
command writeSolution(s,o) ::= if write $\in$ m(s,o)
then
enter read into m(s,o);
fi
```
- readSample
Informal Policy: “Student solutions can be submitted only before downloading any sample solution.” $\Leftrightarrow$ “If the automaton receives an input⟨readSample,(s,o)⟩and the conditions are satisfied, it transitions to a state wheresis denied to submit a solution.”
```
command readSample(s,o) ::= if read$\in$ m(s,o)
then
delete write from m(s,o);
fi
```
3. Initialization
- By model definition: $q_0 =⟨S_0 ,O_0 ,m_0 ⟩$
- For a course with (initially) three students:
- $S_0 =\{sAnn, sBob, sChris\}$
- $O_0 =\{oAnn, oBob, oChris\}$
- $m_0$:
- $m_0(sAnn,oAnn)=\{write\}$
- $m_0(sBob,oBob)=\{write\}$
- $m_0(sChris,oChris)=\{write\}$
- $m_0(s,o)=\varnothing \Leftrightarrow s\not= o$
- Interpretation: “There is a course with three students, each of whom has their own workspace to which she is allowed to submit (write) a solution.”
- Reminder: “For a given security model, is it possible that a subjecteverobtains a specific permission with respect to a specific object?”
- Analysis of Right Proliferation $\rightarrow$ The HRU safety problem.
InputSequences
- “What is the effect of an input in a given state?” $\rightarrow$ asingle state transitionas defined by $\delta$
- “What is the effect of an input sequence in a given state?” $\rightarrow$ a composition ofsequential state transitionsas defined by $\delta*$
> Transitive State Transition Functionδ∗
>
> Let $\sigma\sigma\in\sum^*$ be a sequence of inputs consisting of a single input $\sigma\in\sum\cup\{\epsilon\}$ followed by a sequence $\sigma\in\sum^∗$, where $\epsilon$ denotes an empty input sequence. Then, $\delta∗:Q\times\sum^∗\rightarrow Q$ is defined by
A state q of an HRU model is called HRU safe with respect to a right $r\in R$ iff, beginning with q, there is no sequence of commands that enters r in an ACM cell where it did not exist in q.
According to Tripunitara and Li [2013], this property (Due to more technical details, it’s called simple-safety there.) is defined as:
> HRU Safety
>
> For a state $q=\{S_q,O_q,m_q\}\in Q$ and a right $r\in R$ of an HRU model $⟨Q,\sum,\delta,q_0,R⟩$, the predicate $safe(q,r)$ holds iff
- Theorem 1: See Harrison et al. [1976], reduction to the Halteproblem.
- Theorem 2: We’ll have a closer look at this one ...
- Insights into the operational principles modeled by HRU models
- Demonstrates a method to prove safety property for a particular, given model
- $\rightarrow$ “Proofs teach us how to build things so nothing more needs to be proven.” (W. E. Kühnhauser)
##### Proof of Theorem
- Proof Sketch
1. Find an upper bound for the length of all input sequences with different effects on the protection state w.r.t. safety
If such can be found: $\exists$ a finite number of input sequences with different effects
2. All these inputs can be tested whether they violate safety. This test terminates because:
- each input sequence is finite
- there is only a finite number of relevant sequences
- $\rightarrow$ safety is decidable
Given a mono-operational HRU model.
Let $\sigma_1...\sigma_n$ be any sequence of inputs in $\sum^*$ that violates $safe(q,r)$, and let $p_1...p_n$ be the corresponding sequence of primitives (same length, since mono-operational).
Proposition: For each such sequence, there is a corresponding finite sequence that
- Still violates $safe(q,r)$
- Consists only of enter and two initial create primitives
In other words: For any input sequence,$\exists$ a finite sequence with the same effect.
Proof:
- We construct these finite sequences ...$\rightarrow$
- Transform $\sigma_1...\sigma_n$ into shorter sequences with the same effect:
1. Remove all input operations that contain delete or destroy primitives. The sequence still violates $safe(q,r)$, because conditions of successive commands must still be satisfied (no absence, only presence of rights is checked).
2. Prepend the sequence with an initial create subject $s_{init}$ operation. This won’t change its netto effect, because the new subject isn’t used anywhere.
3. Prune the last create subject s operation and substitute each following reference to s with $s_{init}$. Repeat until allcreate subjectoperations are removed, except from the initialcreate subject sinit.
4. Same as steps 2 and 3 for objects.
5. Remove all redundant enter operations (remember: each matrix cell is a set $\rightarrow$ unique elements).
| enter r1 into m(x2,x5); | enter r1 into m(x2,x5); | enter r1 into m(x2,x5); | enter r1 into $m(s_{init},x5)$; | enter r1 into $m(s_{init},o_{init})$; | enter r1 into $m(s_{init},o_{init})$; |
| enter r2 into m(x2,x5); | enter r2 into m(x2,x5); | enter r2 into m(x2,x5); | enter r2 into $m(s_{init},x5)$; | enter r2 into $m(s_{init},o_{init})$; | enter r2 into $m(s_{init},o_{init})$; |
| enter r1 into m(x7,x5); | enter r1 into m(x7,x5); | enter r1 into m(x7,x5); | enter r1 into $m(s_{init},x5)$; | enter r1 into $m(s_{init},o_{init})$; | - |
| ... | ... | ... | ... | ... | ... |
Observations
- after step 3:
- Except for $s_{init}$, the sequence creates no more subjects
- All rights of the formerly created subjects are accumulated in $s_{init}\rightarrow$ for the evaluation of $safe(q,r)$, nothing has changed:
- This sequence still violates $safe(q,r)$, but its length is restricted to $(|S_q| + 1)(|O_q|+1)|R|+2$ because
- Each enter must enter a new right into a cell
- The number of cells is restricted to $(|S_q| + 1)(|O_q|+1)$
Conclusions from these Theorems
- Dilemma:
- General (unrestricted) HRU models
- have strong expressiveness $\rightarrow$ can model a broad range of AC policies
- are hard to analyze: algorithms and tools for safety analysis
- $\rightarrow$ cannot certainly produce accurate results
- $\rightarrow$ are hard to design for approximative results
- Mono-operational HRU models
- have weak expressiveness $\rightarrow$ goes as far as uselessness: e. g. for modeling Unix creat(can only create files, sockets, IPC, ... that no user process can access!)
- are efficient to analyze: algorithms and tools for safety analysis
- $\rightarrow$ are always guaranteed to terminate
- $\rightarrow$ are straight-forward to design
Consequences:
- Model variants with restricted yet usable expressiveness have been proposed
- Heuristic analysis methods try to provide educated guesses about safety of unrestricted HRU
- Applications: (static) real-time systems, closed embedded systems
Monotonous Mono-conditional HRU Models
- Monotonous (MHRU): no delete or destroy primitives
- Mono-conditional: at most one clause in conditions part (For monotonous bi-conditional models, safety is already undecidable ...)
- safe(q,r) efficiently decidable
- Applications: Archiving/logging systems (where nothing is ever deleted)
Finite Subject Set
- $\forall q\in Q,\exists n\in N: |S_q|\leq n$
- $safe(q,r)$ decidable, but high computational complexity
Fixed STS
- All STS commands are fixed, match particular application domain (e.g. OS access control [Lipton and Snyder, 1977]) $\rightarrow$ no model reusability
- For Lipton and Snyder [1977]: $safe(q,r)$ decidable in linear time (!)
Strong Type System
- Special model that generalizes HRU: Typed Access Matrix (TAM) [Sandhu, 1992]
- $safe(q,r)$ decidable in polynomial time for ternary, acyclic, monotonous variants
- high, though not unrestricted expressiveness in practice
##### (B) Heuristic Analysis Methods
Motivation:
- Restricted model variants: often too weak for real-world applications
- General HRU models: safety property cannot be guaranteed $\rightarrow$ Let’s try to get a piece from both cakes: Heuristically guided safety estimation [Amthor et al., 2013]
Idea:
- State-space exploration by model simulation
- Task of heuristic: generating input sequences (“educated guessing”)
Outline: Two-phase-algorithm to analyze $safe(q_0,r)$:
1. Static phase: Infer knowledge from the model that helps heuristic to make “good” decisions.
- $\rightarrow$ Runtime: polynomial in model size ($q_0 + STS$)
2. Simulation phase: The automaton is implemented and, starting with $q_0$, fed with inputs $\sigma=⟨op,x⟩$
- $\rightarrow$ For each $\sigma$, the heuristic has to decide:
- which operation op to use
- which vector of arguments x to pass
- which $q_i$ to use from the states in $Q$ known so far
- Termination: As soon as $\sigma(q_i,\sigma)$ violates $safe(q_0,r)$.
Goal: Iteratively build up the (possibly infinite!) $Q$ for a model to falsify safety by example (finding a violating, but possible protection state).
Results:
- Termination: Well ... we only have a semi-decidable problem here: It can be guaranteed that a model is unsafe if we terminate. We cannot ever prove the opposite, however! ($\rightarrow$ safety undecidability)
- Performance: A few results
- 2013:Model size 10 000 ≈2215 s
- 2018:Model size 10 000 ≈0,36 s
- 2018:Model size 10 000 000 ≈417 s
Achievements:
- Find typical errors in security policies: Guide their designers, who might know there’s something wrong w. r. t. right proliferation, but not what and why!
- Increase our understanding of unsafety origins: By building clever heuristics, we started to understand how we might design specialized HRU models ($\rightarrow$ fixed STS, type system) that are safety-decidable yet practically (re-) usable [Amthor and Rabe, 2020].
##### Summary HRU Models
Goal
- Analysis of right proliferation in AC models
- Assessing the computational complexity of such analyses
Method
- Combining ACMs and deterministic automata
- Defining $safe(q,r)$ based on this formalism
Conclusions
- Potential right proliferation (privilege escalation): Generally undecidable problem
- $\rightarrow$ HRUmodel family, consisting of application-tailored, safety-decidable variants
- $\rightarrow$ Heuristic analysis methods for practical error-finding
##### The Typed-Access-Matrix Model (TAM)
Goal
- AC model, similar expressiveness to HRU
- $\rightarrow$ can be directly mapped to implementations of an ACM: OS ACLs, DB permission assignment tables
- Better suited for safety analyses: precisely statemodel properties for decidable safety
Idea [Sandhu, 1992]
- Adopted from HRU: subjects, objects, ACM, automaton
- New:leverage the principle of strong typing known from programming
- $\rightarrow$ safety decidability properties relate to type-based restrictions
How it Works:
- Foundation of a TAM model is an HRU model $⟨Q,\sum,\delta,q_0 ,R⟩$, where $Q= 2^S\times 2^O\times M$
- However: $S\subseteq O$, i. e.:
- all subjects can also act as objects (=targets of an access)
- $\rightarrow$ useful for modeling e. g. delegation (“s has the right to grant s′ her read-right”)
- objects in $O\backslash S$: pure objects
- Each $o\in O$ has a type from a type set $T$ assigned through a mapping $type:O\rightarrow T$
> A TAM model is a deterministic automaton $⟨Q,\sum,\delta,q_0 ,T,R⟩$ where
> - $Q= 2^S\times 2^O\times TYPE\times M$ is the state space where $S$ and $O$ are subjects set and objects set as in HRU, where $S\subseteq O$, $TYPE=\{type|type:O\rightarrow T\}$ is a set of possible type functions, $M$ is the set of possible $ACMs$ as in HRU,
> - $\sum=OP\times X$ is the (finite) input alphabet where $OP$ is a set of operations as in HRU, $X=O^k$ is a set of $k$-dimensional vectors of arguments (objects) of these operations,
> - $\delta:Q\times\sum\rightarrow Q$ is the state transition function,
> - $q_0\in Q$ is the initial state,
> - $T$ is a static (finite) set of types,
> - $R$ is a (finite) set of access rights.
State Transition Scheme (STS)
$\delta:Q\times\sum\rightarrow Q$ is defined by a set of specifications:
where
- $q= (S_q,O_q,type_q,m_q)\in Q,op\in OP$
- $r_1,...,r_m\in R$
- $x_{s1},...,x_{sm}\in S_q,x_{o1},...,x_{om}\in Oq\backslash S_q$, and $t_1,...,t_k\in T$ where $s_i$ and $o_i, 1\leq i\leq m$ , are vector indices of the input arguments: $1\leq s_i,o_i\leq k$
- Will control right revocation $\rightarrow$ essence of originator control
- Are not monotonic (consequences ...)
- Summary
- Contributions of ORCON Example
- Owner (“originator”) retains full control over
- Use of her confined objects by third parties $\rightarrow$ transitive right revocation
- Subjects using (or misusing) these objects $\rightarrow$ destruction of these subjects
- Subjects using such objects are confined: cannot forward read information
##### TAM Safety Decidability
Why all this?
- General TAM models (cf. previous definition) $\rightarrow$ safety not decidable (no surprise, since generalization of HRU)
- MTAM:monotonous TAM models; STS without delete or destroy primitives $\rightarrow$ safety decidable if mono-conditional only
- AMTAM:acyclic MTAM models $\rightarrow$ safety decidable, but (most likely) not efficiently: NP-hardproblem
- TAMTAM: ternaryAMTAM models; each STS command requires max. 3 arguments $\rightarrow$ provably same computational power and thus expressive power as AMTAM; safety decidable in polynomial time
##### Acyclic TAM Models
Auxiliary analysis tools for TAM models:
> Parent- and Child-Types
>
> For any operation $op$ with arguments $⟨x_1,t_1⟩,⟨x_2,t_2⟩,...,⟨x_k,t_k⟩$ in an STS of a TAM model, it holds that $t_i, 1\leq i\leq k$
> - is a child type in op if one of its primitives creates a subject or object $x_i$ of type $t_i$,
> - is a parent type in op if none of its primitives creates a subject or object $x_i$ of type $t_i$.
> Type Creation Graph
>
> The type creation graph $TCG=⟨T,E=T\times T⟩$ for the STS of a TAM model is a directed graph with vertex set $T$ and an $edge⟨u,v⟩\in E$ iff $\exists op\in OP:u$ is a parent type in $op\wedge v$ is a child type in op.
Note:In bar,u is both a parent type (because of $s_1$) and a child type (because of $s_2$) $\rightarrow$ hence the loop edge.
Safety Decidability
- We call a TAM model acyclic, iff its TCG is acyclic.
> Theorem [Sandhu, 1992, Theorem 5]
>
> Safety of a ternary, acyclic, monotonous TAM model (TAMTAM) is decidable in polynomial time in the size of $m_0$.
- Crucial property acyclic, intuitively:
- Evolution of the system (protection state transitions) checks both rights in the ACMas well as argument types
- TCG is acyclic $\Rightarrow\exists$ a finite sequence of possible state transitions after which no input tuple with argument types, that were not already considered before, can be found
- One may prove that an algorithm, which tries to expandall possible different follow-up states from $q_0$, may terminate after this finite sequence
- Proof details: SeeSandhu [1992].
Expressive Power of TAMTAM
- MTAM: obviously same expressive power as monotonic HRU (MHRU) $\rightarrow$ cannot model:
- transfer of rights: “take r from ... and in turn grant r to ...”
- countdown rights: “r can only be used n times”
- ORCON example (and many others): allow to ignore non-monotonic command $s$ from STS, e.g. 4.-7., since they
- only remove rights
- are reversible (e. g.: undo 4. by 2.; compensate 7. by 3. where the new subject takes roles of the destroyed one)
- AMTAM: most MTAM STS may be re-written as acyclic(cf. ORCON example)
- TAMTAM: expressive power equivalent to AMTAM
IBAC Model Comparison
- So far: family of IBAC models to describe different ranges of security policies they are able to express(depicted as an Euler diagram):
- Analyze them w. r. t. basic security properties (right proliferation)
- $\rightarrow$ Minimize specification errors
- $\rightarrow$ Minimize implementation errors
- Approach
- Unambiguous policy representation through formal notation
- Prediction and/or verification of mission-critical properties
- Derivation of implementation concepts
- Model Range
- Static models:
- Access control function (ACF): $f:S\times O\times OP\rightarrow \{true,false\}$
- Access control matrix (ACM): $m:S\times O\rightarrow 2^{OP}$
- $\rightarrow$ Static analysis: Which rights are assigned to whom, which (indirect) information flows are possible
- $\rightarrow$ Implementation: Access control lists (ACLs), e.g. in OS, (DB)IS
- Dynamic models:
- ACM plus deterministic automaton $\rightarrow$ Analysis of dynamic behavior: HRU safety
- generally undecidable
- decidable under specific restrictions: monotonous mono-conditional, static, typed, etc.
- identifying and explaining safety-violations, in case such (are assumed to) exists: heuristic analysis algorithms
- Limitations
- IBAC models are fundamental: KISS
- IBAC models provide basic expressiveness only:
- Comparable to “assembler programs for writing AC policies”
- Imagine writing a sophisticated end-user application in assembler:
- reserve and keep track of memory layout and addresses ≈ create and maintain individual rights for thousands of subjects, billions of objects
- display comfortable GUI by writing to the video card framebuffer ≈ specify sophisticated workflows through an HRU STS
- For more application-oriented policy semantics:
- Large information systems: many users, many databases, files, ... $\rightarrow$ Scalability problem
- Access decisions not just based on subjects, objects, and operations $\rightarrow$ Abstraction problem
$\rightarrow$ “New” paradigm (early–mid 90s): Role-based Access Control
#### Roles-based Access Control Models (RBAC)
> Self-Study Task
>
> Have a look at the syscall API of Linux as a typical file server operating system. Roughly count the number of operations and estimate their average number of arguments based on a few samples. Then try to estimate the average number of files that each user keeps in her home directory. A good sample is your own user directory, which you can count (including subdirectories) as follows: `find ~ -type f | wc -l`
>
> If 200 employees of a medium-sized company have user accounts:
> - How many ACLs must be saved to encode the IBAC policy of this server as a classical ACM?
> - If each ACL take 12 bits, how big is the resulting storage overhead in total?
> - If you had to heuristically analyze safety of this policy: how many different inputs would you have to simulate in the worst case just for the first state transition?
Problems of IBAC Models:
- Scalability w.r.t. the number of controlled entities
- Level of abstraction: System-oriented policy semantics (processes, files, databases, ...) instead of problem-oriented (management levels, user accounts, quota, ...)
Goals of RBAC:
- Solving these problems results in smaller modeling effort results in smaller chance of human errors made in the process:
- Improved scalability and manageability
- Improved, application-oriented semantics: roles≈functions in organizations
