This chapter is from the book
This chapter is from the book
This chapter is from the book
1.13 The Multi-level Model of Security
Computer security has become sufficiently important that it was inevitable that governments would decide they needed to "do something about it". And when governments want to know something about security, they turn to the experts: the military. And they develop standards and measurement tools by which security can be measured that are unbiased so as not to favor any one organization. And they mandate that anyone they can influence buy products meeting those standards.
The problem is that secure is not as simple to define as, say, flame-retardant. The security threats in different environments are very different, as are the best ways to counter them. The military has traditionally focussed on keeping their data secret (and learning the secrets of the other side). They are less concerned (though probably shouldn't be) about data getting corrupted or forged. In a paper world, forgeries are so difficult and so likely to expose the spies placing them that this threat takes a back seat. In the computerized environment, modification or corruption of data is a more likely threat.
1.13.1 Mandatory (Nondiscretionary) Access Controls
O negligence! ... what cross devil made me put this main secret in the packet I sent the king? Is there no way to cure this? No new device to beat this from his brains?
—Shakespeare's King Henry VIII, act 3, scene 2
Discretionary means that someone who owns a resource can make a decision as to who is allowed to use (access) it. Nondiscretionary access controls enforce a policy where users might be allowed to use information themselves but might not be allowed to make a copy of it available to someone else. Strict rules are automatically enforced about who is allowed access to certain resources based on the attributes of the resource, and even the owners of the resources cannot change those attributes. The analogy in the paper world is that you might be given a book full of confidential information, but you are not allowed to take the book out of the building. In the military, information often has a security classification, and just because you have access to secret information does not mean you can forward it as you see fit. Only someone with the proper clearance can see it, and clearance levels are decided by a separate organization based on background investigations—not based on whether someone seems like a nice guy at lunch.
The basic philosophy behind discretionary controls is that the users and the programs they run are good guys, and it is up to the operating system to trust them and protect each user from outsiders and other users. The basic philosophy behind nondiscretionary controls is that users are careless and the programs they run can't be presumed to be carrying out their wishes. The system must be ever vigilant to prevent the users from accidentally or intentionally giving information to someone who shouldn't have it. Careless users might accidentally type the wrong file name when including a file in a mail message, or might leave a message world-readable. The concept is to confine information within a security perimeter, and thus not allow any information to move from a more secure environment to a less secure environment. A secure system would have both discretionary and nondiscretionary access controls, with the latter serving as a backup mechanism with less granularity.
Now of course, if you allow the users out of the building, there is an avenue for information to leak out of a secure environment, since a user can remember the information and tell someone once the user gets out of the security perimeter. There really is no way for a computer system to prevent that. But the designers wanted to ensure that no Trojan horse in software could transmit any information out of the perimeter, that nothing a user did inadvertently could leak information, and that users couldn't spirit out larger amounts of information than they could memorize.
1.13.2 Levels of Security
What does it mean for something to be "more sensitive" than something else? We will use a somewhat simplified description of the U.S. Department of Defense (DoD) definitions of levels of security as an example. It is a reasonably general model and similar to what is done in other contexts. It is sufficient to understand the security mechanisms we'll describe.
The security label of something consists of two components:
A security level (also known as classification), which might be an integer in some range, but in the U.S. DoD consists of one of the four ratings unclassified, confidential, secret, and top secret, where _unclassified < confidential < secret < top secret.
A set of zero or more categories (also known as compartments), which describe kinds of information. For instance, the name CRYPTO might mean information about cryptographic algorithms, INTEL might mean information about military intelligence, COMSEC might mean information about communications security, or NUCLEAR might mean information about types of families.
Documents (or computer files) are marked with a security label saying how sensitive the information is, and people are issued security clearances according to how trustworthy they are perceived to be and what information they have demonstrated a "need to know."
A clearance might therefore be (SECRET;{COMSEC,CRYPTO}), which would indicate someone who was allowed to know information classified unclassified, confidential, or secret (but not top secret) dealing with cryptographic algorithms or communications security.
Given two security labels, (X, S1) and (Y, S2 ), (X, S1) is defined as being "at least as sensitive as" (Y, S2 ) iff X σ Y and S2 Õ S1. For example,
(TOP SECRET, {CRYPTO, COMSEC}) > (SECRET, {CRYPTO})
where ">" means "more sensitive than".
It is possible for two labels to be incomparable in the sense that neither is more sensitive than the other. For example, neither of the following are comparable to each other:
(TOP SECRET, {CRYPTO, COMSEC})
(SECRET, {NUCLEAR, CRYPTO})
1.13.3 Mandatory Access Control Rules
Every person, process, and piece of information has a security label. A person cannot run a process with a label higher than the person's label, but may run one with a lower label. Information is only allowed to be read by a process that has at least as high a rating as the information. The terminology used for having a process read something with a higher rating than the process is read-up. Read-up is illegal and must be prevented. A process cannot write a piece of information with a rating lower than the process's rating. The terminology used for a process writing something with a lower rating than the process is write-down. Write-down is illegal and must be prevented.
The rules are:
A human can only run a process that has a security label below or equal to that of the human's label.
A process can only read information marked with a security label below or equal to that of the process.
A process can only write information marked with a security label above or equal to that of the process. Note that if a process writes information marked with a security label above that of the process, the process can't subsequently read that information.
The prevention of read-up and write-down is the central idea behind mandatory access controls. The concepts of confinement within a security perimeter and a generalized hierarchy of security classes were given a mathematical basis by Bell and La Padula in 1973 [BELL74]. There is significant complexity associated with the details of actually making them work. There has been significant subsequent research on more complex models that capture both the trustworthiness and the confidentiality of data and programs.
1.13.4 Covert Channels
A covert channel is a method for a Trojan horse to circumvent the automatic confinement of information within a security perimeter. Let's assume an operating system has enforced the rules in the previous section. Let's assume also that a bad guy has successfully tricked someone with a TOP SECRET clearance into running a program with a Trojan horse. The program has access to some sensitive data, and wants to pass the data to the bad guy. We're assuming the operating system prevents the process from doing this straightforwardly, but there are diabolical methods that theoretically could be employed to get information out.
The Trojan horse program cannot directly pass data, but all it needs is for there to be anything it can do that can be detected by something outside the security perimeter. As long as information can be passed one bit at a time, anything can be transmitted, given enough time.
One kind of covert channel is a timing channel. The Trojan horse program alternately loops and waits, in cycles of, say, one minute per bit. When the next bit is a 1, the program loops for one minute. When the next bit is a 0, the program waits for a minute. The bad guy's program running on the same computer but without access to the sensitive data constantly tests the loading of the system. If the system is sluggish, its conspirator inside the perimeter is looping, and therefore transmitting a 1. Otherwise, the conspirator is waiting, and therefore transmitting a 0.
This assumes those two processes are the only ones running on the machine. What happens if there are other processes running and stopping at seemingly random times (from the point of view of the program trying to read the covert channel)? That introduces noise into the channel. But communications people can deal with a noisy channel; it just lowers the potential bandwidth, depending on the signal to noise ratio.
Another kind of covert channel called a storage channel involves the use of shared resources other than processor cycles. For instance, suppose there were a queue of finite size, say the print queue. The Trojan horse program could fill the queue to transmit a 1, and delete some jobs to transmit a 0. The covert channel reader would attempt to print something and note whether the request was accepted. Other possible shared resources that might be exploited for passing information include physical memory, disk space, network ports, and I/O buffers.
Yet another example depends on how clever the operating system is about not divulging information in error messages. For instance, suppose the operating system says file does not exist when a file really does not exist, but says insufficient privilege for requested operation when the file does exist, but inside a security perimeter off limits to the process requesting to read the file. Then the Trojan horse can alternately create and delete a file of some name known to the other process. The conspirator process periodically attempts to read the file and uses the information about which error message it gets to determine the setting of the next bit of information.
There is no general way to prevent all covert channels. Instead, people imagine all the different ways they can think of, and specifically attempt to plug those holes. For instance, the timing channel can be eliminated by giving each security level a fixed percentage of the processor cycles. This is wasteful and impractical in general, because there can be an immense number of distinct classifications (in our model of (one of four levels, {categories}), the number of possible security perimeters is 4 2n, where n is the number of categories).
Most covert channels have very low bandwidth. In many cases, instead of attempting to eliminate a covert channel, it is more practical to introduce enough noise into the system so that the bandwidth becomes too low to be useful to an enemy. It's also possible to look for jobs that appear to be attempting to exploit covert channels (a job that alternately submitted enough print jobs to fill the queue and then deleted them would be suspicious indeed if someone knew to watch). If the bandwidth is low and the secret data is large, and knowing only a small subset of the secret data is not of much use to an enemy, the threat is minimized.
How much secret data must be leaked before serious damage is done can vary considerably. For example, assume there is a file with 100 megabytes of secret data. The file has been transmitted, encrypted, on an insecure network. The enemy therefore has the ciphertext, but the cryptographic algorithm used makes it impossible for the enemy to decrypt the data without knowing the key. A Trojan horse with access to the file and a covert channel with a bandwidth of 1 bit every 10 seconds would require 250 years to leak the data (by which time it's hard to believe the divulging of the information could be damaging to anyone. However, if the Trojan horse had access to the 56-bit key, it could leak that information across the covert channel in less than 10 minutes. That information would allow the enemy to decrypt the 100-megabyte file. For this reason, many secure systems go to great pains to keep cryptographic keys out of the hands of the programs that use them.
1.13.5 The Orange Book
The National Computer Security Center (NCSC), an agency of the U.S. government, has published an official standard called "Trusted Computer System Evaluation Criteria", universally known as "the Orange Book" (guess what color the cover is). The Orange Book defines a series of ratings a computer system can have based on its security features and the care that went into its design, documentation, and testing. This rating system is intended to give government agencies and commercial enterprises an objective assessment of a system's security and to goad computer manufacturers into placing more emphasis on security.
The official categories are D, C1, C2, B1, B2, B3, and A1, which range from least secure to most secure. In reality, of course, there is no way to place all the possible properties in a linear scale. Different threats are more or less important in different environments. The authors of the Orange Book made an attempt to linearize these concerns given their priorities. But the results can be misleading. An otherwise A1 system that is missing some single feature might have a D rating. Systems not designed with the Orange Book in mind are likely to get low ratings even if they are in fact very secure.
The other problem with the Orange Book rating scheme is that the designers focused on the security priorities of military security people—keeping data secret. A rating of B1 or better requires implementation of multi-level security and mandatory access controls. In the commercial world, data integrity is at least as important as data confidentiality. Mandatory access controls, even if available, are not suitable for most commercial environments because they make some of the most common operations, such as having a highly privileged user send mail to an unprivileged user, very cumbersome.
Mandatory access controls do not by themselves protect the system from infection by viruses. Mandatory access controls allow write-up, so if some unprivileged account became infected by having someone carelessly run, say, a game program loaded from a bulletin board, the virus could spread to more secure areas. Ironically, if it was a very secure area that first got infected, the mandatory access control features would prevent the infection from spreading to the less secure environments.
The following is a summary of what properties a system must have to qualify for each rating.
D – Minimal Protection. This simply means the system did not qualify for any of the higher ratings; it might actually be very secure. No system is ever going to brag about the fact that it was awarded a D rating.
C1 – Discretionary Security Protection. The requirements at this level correspond roughly to what one might expect from a classic timesharing system. It requires:
The operating system must prevent unprivileged user programs from overwriting critical portions of its memory. (Note that many PC operating systems do not satisfy this condition.)
Resources must be protected with access controls. Those access controls need not be sophisticated; classic owner/group/world controls would be sufficient.
The system must authenticate users by a password or some similar mechanism, and the password database must be protected so that it cannot be accessed by unauthorized users.
There are additional requirements around testing and documentation, which become more detailed at each successive rating.
C2 – Controlled Access Protection. This level corresponds roughly to a timesharing system where security is an important concern but users are responsible for their own fates; an example might be a commercial timesharing system. The additional requirements (over those required for C1) for a C2 rating are:
access control at a per user granularity—It must be possible to permit access to any selected subset of the user community, probably via ACLs. An ACL is a data structure attached to a resource that specifies the resource's authorized users. C2 does not explicitly require ACLs, but they are the most convenient way to provide the granularity of protection that C2 requires.
clearing of allocated memory—The operating system must ensure that freshly allocated disk space and memory does not contain "left-over" data deleted by some previous user. It can do that by writing to the space or by requiring processes to write to the space before they can read it.
auditing—The operating system must be capable of recording security-relevant events, including authentication and object access. The audit log must be protected from tampering and must record date, time, user, object, and event. Auditing must be selective based on user and object.
It is reasonable to expect that C2-rateable systems will become ubiquitous, since they contain features that are commonly desired and do not represent an unacceptable overhead. It is somewhat surprising that such systems are not the norm.
B1 – Labeled Security Protection. Additional requirements at this level are essentially those required to implement Mandatory Access Controls for secrecy (not integrity) except that little attention is given to covert channels. Requirements for B1 above those for C2 include:
Security Labels: Sensitivity labels must be maintained for all users, processes, and files, and read-up and write-down must be prevented by the operating system.
Attached devices must either themselves be labeled as accepting only a single level of information, or they must accept and know how to process security labels.
Attached printers must have a mechanism for ensuring that there is a human-readable sensitivity label printed on the top and bottom of each page corresponding to the sensitivity label of the information being printed. The operating system must enforce this correspondence.
B2 – Structured Protection. Beyond B1, there are few new features introduced; rather, the operating system must be structured to greater levels of assurance that it behaves correctly (i.e., has no bugs). Additional requirements for B2 include:
trusted path to user—There must be some mechanism to allow a user at a terminal to reliably distinguish between talking to the legitimate operating system and talking to a Trojan horse password-capturing program.
security level changes—A terminal user must be notified when any process started by that user changes its security level.
security kernel—The operating system must be structured so that only a minimal portion of it is security-sensitive, i.e., that bugs in the bulk of the O/S cannot cause sensitive data to leak. This is typically done by running the bulk of the O/S in the processor's user mode and having a secure-kernel mini-O/S which enforces the mandatory access controls.
Covert channels must be identified and their bandwidth estimated, but there is no requirement that they be eliminated.
Strict procedures must be used in the maintenance of the security-sensitive portion of the operating system. For instance, anyone modifying any portion must document what they changed, when they changed it, and why, and some set of other people should compare the updated section with the previous version.
B3 – Security Domains. Additional requirements for B3 mostly involve greater assurance that the operating system will not have bugs that might allow something to circumvent mandatory access controls. Additional requirements include:
ACLs must be able to explicitly deny access to named individuals even if they are members of groups that are otherwise allowed access. It is only at this level that ACLs must be able to separately enforce modes of access (i.e., read vs. write) to a file.
active audit—There must be mechanisms to detect selected audited events or thresholds of audited events and immediately trigger notification of a security administrator.
secure crashing—The system must ensure that the crashing and restarting of the system introduces no security policy violations.
A1 – Verified Design. There are no additional features in an A1 system over a B3 system. Rather, there are formal procedures for the analysis of the design of the system and more rigorous controls on its implementation.
1.13.6 Successors to the Orange Book
Gee, I wish we had one of them Doomsday machines
—Turgidson, in Dr. Strangelove
Governments are rarely willing to adopt one another's ideas, especially if they didn't contribute. They would rather develop their own. The publication of the Orange Book in 1983 set off a series of efforts in various countries to come up with their own standards and classifications. These efforts eventually merged in 1990 into a single non-U.S. standard called ITSEC, followed by a reconciliation with the U.S. and the development of a single worldwide standard called the Common Criteria in 1994. Version 2.1 of the Common Criteria became an international standard in 1999.
The details of the various rating systems are all different, so passing the bureaucratic hurdles to qualify for a rating under one system would not be much of a head start toward getting an equivalent rating in another (much as different countries don't recognize each other's credentials for practicing medicine). The following table is an oversimplification that we hope won't be too offensive to the advocates of these rating systems, but it does allow the novice to judge what is being claimed about a system. In a partial acknowledgment of the multifaceted nature of security, many of the systems gave two ratings: one for the features provided and one for the degree of assurance that the system implements those features correctly. In practice, like the artistic and technical merit scores at the Olympics, the two scores tend to be closely correlated.
|
TCSEC (Orange Book) |
German (Green Book) |
British CLEF |
ITSEC |
Common Criteria |
|
D |
Q0 |
|
E0 EAL 0 |
|
|
C1 |
Q1/F1 |
L1 |
C1/E1 |
EAL 2 |
|
C2 |
Q2/F2 |
L2 |
C2/E2 |
CAPP/EAL 3 |
|
B1 |
Q3/F3 |
L3 |
B1/E3 |
CSPP/EAL 4 |
|
B2 |
Q4/F4 |
L4 |
B2/E4 |
EAL 5 |
|
B3 |
Q5/F5 |
L5 |
B3/E5 |
EAL 6 |
|
A1 |
Q6/F5 |
L6 |
B3/E6 |
EAL 7 |