Abstract

Deception is a classic technique useful for military operations.� With information systems around the world under frequent attack every day, it is appropriate to consider analogies from conventional warfare, and deception has historically been powerful as both a tactic and a strategy.� We here systematically enumerate and rank the available deception options for information systems, both offensively and defensively.� We then consider how defensive deceptions can be packaged within "generic excuses" that will more convincing to an attacker than isolated refusals to obey commands.� We describe how the selection of the best generic excuses and excuse application times can be formulated with probabilities as an optimization problem and solved.� Our theory lends itself well to computer implementation and we provide several examples.

This paper appeared in the IEEE First Symposium on Multi-Agent Security and Survivability, Philadelphia, August 2004.

Keywords

deception, information security, security and privacy protection, human communications, lying, multiagent systems, Bayesian inference, excuses

INTRODUCTION

Deception is an important two-agent psychological phenomenon with many applications to information security.� Deception can occur offensively, as when attackers try to fool our information systems into giving away secrets or destroying themselves, or it can occur defensively, as when a computer pretends by exaggerated processing delays to succumb to a denial-of-service attack so the attacker goes away.� Little systematic analysis has been paid to the concept of deception in information systems, however.� Most writers assume methods of deception are intuitively obvious or comprise only a few categories.� But there are many possible kinds of deception, some quite subtle.� And deceptions can be automated.� So it is important in defense of our information systems that we develop a better theory of deception and ways to translate it into security practices.

Our main interest is in using deception to defend information systems, as a second line of defense when access controls have been breached on those systems.� Over a thousand attack methods a year are being reported to the Carnegie-Mellon Computer Emergency Response Team and they are increasing in sophistication [21].� Deception can be highly effective against these attacks since information systems are thought to be totally honest.� For instance, when asked to download a suspicious file, a deceptive operating system could say the network isn't working, or could pretend to download it, or could download it and delete later when the attacker isn't checking.� It may take an attacker a while to realize they are being deceived.� This time may be critical in denial-of-service attacks and "information warfare" when the attacker depends on surprise.� But we need a framework for proposing deceptive practices in defense of information systems because there are so many options to consider.� Such a framework must also model deceptions that attackers use so we can find good counter-deceptions to foil them.� We could implement defensive deceptions in modified operating systems.

Is this kind of deception ethical?� The ethics of deception have been much debated over the centuries [6].� But there are cases in which it is generally justified, as when the costs of no deception are much more severe than the costs of deception.� This can be argued is true for a computer system where to allow an attack can make it inoperable with worthless data ? provided our deceptions are carefully enough triggered to rarely catch legitimate users.� So we will proceed here on the assumption that a broad range of deceptions can be ethical in defending our information systems against serious attacks.

What about automated attacks and script-following ("script kiddie") attacks?� Deception would seem to need a thinking person to fool.� We believe it more valuable for security analysis to focus on assaults on important computer systems where attackers are thinking carefully about what they are doing.� Nonetheless, automated and semi-automated attacks are easily tripped up by small surprises like unexpectedly absent resources.� Most of the deceptions we will enumerate here are surprises in some sense and thus will also foil automated and semi-automated attacks.

To illustrate the issues, consider a Web service we have implemented.� If users to our "demos" site click on a link saying "Secret file directory access" they see a typical-looking file directory (Figure 1, upper right).� If they click on "org.htm" they see an organization chart with unexplained acronyms (lower right).� If they examine several of these files, they see what appears to be encrypted data (upper left).� If they try to open "/root/sigs/ bio_weber/~thornton/13Hands", they get the message "Error 401: No detailed message"; if they try "/root/sigs/bio_weber/leaver/ 1000m/panel.cry", they get� "You are not authorized to view this page"; and if they try "/root/sigs/bio_weber/ module6.rzp", the program waits ten minutes and then says the page is too large to display.� Other images shown for this directory include pictures of someone under "bio_weber" (apparently Mr. Weber), wood grain, a blurry picture of someone holding what appears to be a toy blimp, PowerPoint slides about wireless LANs, an Iraqi missile launcher, a strange experimental vehicle being loaded by crane on a ship (lower left in Figure 1), an explosion cloud, and a graphic labeled "Mishap Plans" (under /root/sigs/bio web/leaver/1000m/Images/Now is the time_ files/avsafety/gouge.� It looks like information about some military organization where Mr. Weber is the founder and Thornton and Leaver are current employees engaged in something involving sinister secret networked weapons that gouge things, whose shortly imminent use ("now is the time") is intended to look accidental (a "mishap").

This is a prototype of a new kind of "honeypot" designed to attract spies and waste their time.� Information derived from real data is put together randomly to encourage the human tendency to see patterns where none exist.� The behavior is generated by a Java servlet; there are no files and directories except for one big file of images, captions, and Web page names extracted by a crawler from easily accessible military Web pages.� The page names are used to randomly create reasonable-sounding directory and subdirectory names, with random extensions added to the file names.� Those with "htm" extensions mostly pull related image-caption pairs from the database.� Those with other extensions create different kinds of random data (from a pseudorandom seed that is a function of the file size to ensure consistency).� The system exaggerates loading delays of the fake files, and if the imaginary file is listed as especially large, the user is told the file is too large to load after a long wait.� Randomly 10% of the time the user is told that they do not have authorization to view a file (with two kinds of error messages).� All in all, the site is designed to make a spy think there are secrets here and to see nonexistent connections, a classic way to neutralize spying agencies as was done to U.S. CIA counterintelligence during the 1970s.� It could also frustrate legitimate users but legitimate users should not be looking for secrets.� This deceptive software was carefully crafted; a question is whether design of similar deceptive software could be automated.

Surprisingly, not much work has addressed the idea of deception in defense of computer systems.� Some good work has developed honeypot computers and "honeynets", computer systems and networks whose purpose is to entrap attackers to obtain information about their attack methods [18]; these are a simple passive form of deception.� [1] suggests a general approach to planning defenses automatically in adversarial situations, including classic military tactics like deception; [8] suggests ways to encourage maladaptive behavior in enemy organizations.� [10] and [19] provide ideas for modeling attacks on computer systems, the former suggesting some interesting ideas about deception in defense.� Work on multi-agent systems is starting to examine societies of deceptive agents [3, 9, 30].

Figure 1: Example screens from the honeypot.

A TAXONOMY OF DECEPTION METHODS IN ATTACKS ON INFORMATION SYSTEMS

[4] identifies six basic methods of deception: masking, repackaging, dazzling, mimicking, inventing, and decoying.�� These are good at identifying categories of effect but do not explain how deception occurs.� [31] suggests nine basic "misperception" categories: pattern, players, intention, payoff, place, time, strength, style, and channel.� These are a good start at a taxonomy but fail to incorporate important semantic distinctions developed over the years in artificial intelligence.� Work in military science often refers to deception: [13] proposes nine methods of military deception but these are specific to military operations; [16] provides six useful principles of military deception, but no theory.� [17] provides some useful pieces of a theory, but the concern is analysis of deception rather than generation of deception.

A deeper theory of deception can be developed from semantic cases of computational linguistics [14].� Every action can be associated with other concepts called semantic cases, a generalization of syntactic cases in language.� Various cases have been proposed, of which [11] is the most comprehensive we have seen.� To their 28 we add four more, the upward type-supertype and part-whole links and two speech-act conditions [2], to get 32 altogether:

� essence: supertype (generalization of the action type) and whole (of which the action is a part)

� participant: agent (the person who initiates the action), object (what the action is done to), beneficiary (the person who benefits), experiences (the psychological feature associated with the action), instrument (some thing that helps accomplish the action),� accompaniment (additional object associated with the action) and recipient (the person who receives the action)

� space: location-at, location-from, location-to, location-through (through which the action occurs), direction (of the action), and orientation (in a metric space)

� time: frequency (of occurrence), time-at, time-from, time-to, and time-through

� causality: cause, contradiction (what this action contradicts if anything), effect, and purpose

� speech-act theory: precondition (on the action) and ability (of the agent to perform the action)

� quality: content (type of the action object), manner (the way in which the action is done), material (the atomic units out of which the action is composed), measure (the quantity associated with the action), order (of events), and value (the data transmitted by the action)

Our claim is that deception operates on an action to change its perceived case values.� For instance, an attacker masquerading as the administrator of a computer system is concealing the agent case and purpose case associated with their actions on that system.� For the fake-directory prototype that we built, we use deception in object (it isn't a real directory interface), "time-through" (some responses are deliberately delayed), cause (it lies about files being too big to load), preconditions (it lies about necessary authorization), and effect (it lies about existence of files).

Not all the cases make sense for the interaction between an attacker and the software of a computer system.� We can rule out beneficiary (the attacker is the assumed beneficiary of all activity), experiences (psychological states don't matter in giving commands and obeying them), recipient (there are only two agents), "location-at" (you can't "inhabit" cyberspace), orientation (there is no coordinate system in cyberspace), "time-from" (attacks happen anytime), "time-to" (attacks happen anytime), contradiction (commands don't include comparisons), manner (there's only one way commands execute besides in duration), material (everything is bits in cyberspace), and order (the order of commands or files rarely can be varied).

In general, offensive opportunities for deception are as frequent as defensive opportunities, but appropriate methods differ.� For instance, the instrument case is associated with offensive deceptions in cyberspace since the attacker can attack with a variety of methods: buffer overflows, viruses, backdoors, stolen passwords, etc.� The defender has few defensive instruments and must focus on foiling the attacker's.� In contrast, "time-through" is primarily associated with defensive deceptions in cyberspace since the defending system controls the time it takes to respond to a command, and is little affected by the time it takes an attacker to issue commands to it.� But "object" can be associated with both offense and defense because attackers can choose to attack little-defended targets like unused ports, while defenders can substitute low-value targets like honeypots for high-value targets that the attacker thinks they are compromising.� Table 1 summarizes the suitability of the remaining 21 deception methods as we judge them on general principles.� 10 indicates the most suitable, and 0 indicates unsuitable.� These numbers were obtained by expert-systems methodology and could be refined by surveys of users or deception experiments.� This table provides a wide-ranging menu of choices for deception planners.

Table 1 specializes an artificial-intelligence planning-based approach in which deceptions at a state can be represented as sets of atomic "ploys" of three kinds: Deletion of facts from the state, addition of facts to the state, or change of facts in the state.� For instance, if we want to prevent an attacker from downloading a suspicious file, giving a "network is down" excuse amounts to an addition of a network-status fact to the state and deletion of the expected download from the state.� But this is clearly too limited a model since the manner in which information is conveyed by a deceiver is just as important as the content of the deception.� For instance, a command can be refused by a system, to the same effect, by one of these "ploy presentation tactics":

� Pretend that the attacker's command is not legal syntax.

� Give a false excuse to the attacker why the command cannot be done:

� Say that a necessary resource (network, file system, directory, etc.) is unavailable.

� Say that authorization is required for a necessary resource.

� Claim this is the wrong command for the resource.

� Lie that the command has been done.

� Obey the command now, but undo it later when the attacker is not looking.

� Take such a long time to do the command that the attacker gives up.

� Distract the attacker with so much extraneous information and so many requests that they give up.

� Secretly transfer the attacker to a "safe sandbox" where their commands will be harmless.

� Secretly transfer "spyware" back to the attacker's machine so you can trace them and stop them.

Choosing the ploy presentation tactic will be another key part of a deception strategy.

Table 1: Evaluation of deception methods in cyberspace.

Deception method	Suitability for offense in information systems, with general example	Suitability for defense in information systems, with general example
supertype	6 (pretend attack is something else)	0
whole	8 (conceal attack in a common sequence of commands)	0
agent	4 (pretend attacker is legitimate user or is standard software)	0
object	8 (attack unexpected software or feature of a system)	5 (camouflage key targets or make them look unimportant, or disguise software as different software)
instrument	7 (attack with a surprising tool)	0
accompaniment	9 (a Trojan horse installed on a system)	6 (software with a Trojan horse that is sent to attacker)
location-from	5 (attack from a surprise site)	2 (try to frighten attacker with false messages from authorities)
location-to	3 (attack an unexpected site or port if there are any)	6 (transfer control to a safer machine, as on a honeynet)
location-through	3 (attack through another site)	0
direction	2 (attack backward to site of� a user)	4 (transfer Trojan horses back to attacker)
frequency	10 (swamp a resource with tasks)	8 (swamp attacker with messages or requests)
time-at	5 (put false times in event records)	2 (associate false times with files)
time-through	1 (delay during attack to make it look as if attack was aborted)	8 (delay in processing commands)
cause	1 (doesn't matter much)	9 (lie that you can't do something, or do something not asked for)
effect	3 (lie as to what a command does)	10 (lie as to what a command did)
precondition	5 (give impossible commands)	8 (give false excuses for being unable to do something)
ability	2 (pretend to be an inept attacker or have inept attack tools)	5 (pretend to be an inept defender or have easy-to-subvert software)
content	6 (redefine executables; give false file-type information)	7 (redefine executables; give false file-type information)
measure	5 (send data too large to easily handle)	7 (send data too large or requests too hard to attacker)
value	3 (give arguments to commands that have unexpected consequences)	9 (systematically misunderstand attacker commands)

MODELING TRUST IN DECEPTIONS

Our goal is to provide "counterdeception" for defending information systems, planning to foil an attacker's deceptions with deceptions of our own.� (Attackers could use an analogous approach to improve their attacks.)� Besides the intrinsic suitability of a deception, we must also consider how likely it fools an attacker.

Our estimation that a deception succeeds assumes it is a monotonically decreasing function of the a priori probability of the deceptive action occurring without deception.� Humans who are trained to detect lies look for inconsistencies and improbabilities [15], a form of this principle.� Probabilities can be estimated from statistics of how often events occurred in a normally operating computer system, e.g. how many times the network connection was down on the average.� We can use a sigmoid-shaped function as in neural-network methods for similar situations, a function that is near-flat at both ends but increases from 0 to 1 in the middle of its range.� [12] and [3] provide interesting alternative models.

The Strength of Attacker Hypotheses

We need a model of "suspiciousness" or lack of trust by an attacker in an information system, since this is what we want to manage carefully when we are deceiving an attacker.� Some modern research in sociology on the concept of trust provides analysis that we can use.� [27] claims "trust is a bet about the future contingent actions of others."� Let us represent it as trust(A,D,O,C) which means the probability that attacker A believes defender D (the computer system) will perform order O under condition C.

Though [27] did not formulate this mathematically, this approach is compatible with� Bayes' Rule, which says p(T|E) = p(E|T) p(T) / p(E) where T is the act of trust is subsequently justified by events for a particular attacker A, defender D, order O, and conditions C, and E is some piece of evidence that D provides as to their trustworthiness (like obeying the last command promptly).� When we have several pieces of evidence E1, E2, etc. the formula becomes:

p(T|E1&E2&E3&...) = p(E1&E2&E3&...|T) p(T)�� / p(E1&E2&E3&...)

If the evidence arrives sequentially, we can write this as:

p(T|E1&E2&E3&...) = p(T)� (p(E1|T)/ p(E1))� (p(E2|T&E1)/p(E2|E1)) (p(E3|T&E1&E2)/p(E3|E2&E1))...

When some of the evidence is independent, we can eliminate some of the conjuncts; if all the evidence is independent, we get the "Naive Bayes" formula [32]:

(p(E1|T)/p(E1)) (p(E2|T)/p(E2)) (p(E3|T)/p(E3) ... p(T)

These last two forms conveniently separate each piece of evidence.� Each of these ratios can be greater than 1 (for evidence increasing trust, such as the compliance of D with commands) or less than 1 (for evidence decreasing trust, such as refusal of D to comply with commands).

The main difficulty in applying this formula is not so much that independence does not apply (the formula has been shown to be a useful ranking method in many applications when independence clearly does not hold), but that we rarely have sufficient statistics on the exact situation of interest.� If A has not dealt with D before, or A has not dealt with D in regard to either O or C, A must estimate probabilities.� [9] notes this in modeling deception and tries a different approach, but we do not see this is necessary since one can use the results of a running intrusion-detection system if enough data has been collected about a user, and otherwise "statistical inheritance" with known marginal statistics on generalizations D' of D, O' of O, and C' of C.� If there is more than one such D', O', or C' we need to take a weighted average of their predictions, where typically the weight is the inverse of the corresponding set size (the number of things in the real world to which the class corresponds, since larger classes are further in properties from the target class).� Such reasoning can be fallacious, as for instance when we have had one bad experience with a Microsoft product and we generalize to all Microsoft products, or when we have had one bad experience with a person of one ethnic group and we generalize to all members of that group.� Nonetheless, such reasoning is essential for everyday situations.� So we can assume D is a typical Internet site, or O is a typical network-using command, and use trust in all sites or all network commands for our estimate.� [27] enumerates a comprehensive set of such inferences; for cyberspace, these should include influences from what attacker A hears from other hackers or the general trust level of the hacker "society".

Generic Excuses for Deception

There are several kinds of trust, however, because deception is not the only hypothesis that an attacker may entertain as we attempt to deceive them.� A well-known psychological trait of humans is their ability to see patterns.� So rather than just trying to conceal a deception, it is usually easier to convince an attacker of some different explanation of� events.� Magicians often try to induce such explanations in audience, like "supernatural visitation", "psychic powers", or "new laws of physics", also called "mock explanations" [23].� Some good such "generic excuses" in responding to information-systems attacks are:

� The system, file system, or network has crashed.

� Attacker actions "broke" something on the system or network so it is not working.

� Testing is being done on the system or network.

� Particular software on the system or network is new and not fully debugged.

� Network communication defaults are incorrect.

� Security policy or limits have recently changed on the system or network, affecting user capabilities.

� The system or network has been "hacked" and compromised.

� A practical joker [26] is deliberately trying to irritate users.

To these we must add the generic excuse of "deception", the one hypothesis we do not want the attacker to entertain.� We can assign a priori probabilities to the generic excuses.� These will vary with the system and its setting, but let us assume here typical values with "network down" of relative likelihood 0.10, "testing" 0.10, "bugs" 0.08, "communications faulty" 0.06, "policy change" 0.04, "hacked" 0.01, "practical joker" 0.01, and "deception" 0.01.� Generic excuses should only be given when we are reasonably certain that the user is an attacker.; see below under "The Suspiciousness of the Attacker" and "The Penalty for Deceiving Non-Attackers."

As an illustration, suppose an attacker of a computer system tries to download a file from their home site and gets the error message "no network response"; then they manually modify an operating-system executable and get the message "error in file routines".� Suppose the a priori probabilities of the message "no network response" is 0.02 and "error in file routines" are 0.06.� Suppose the probability of a network error occurring when the network is down is 0.9, the probability of file error when there are bugs in the system is 0.8, and the probabilities of vice versa are 0.01 each.� Suppose the probability of any message given deception is 0.1.� Suppose the message "no network response" occurs 20% of the time when there are network problems, and "error in file routines" appears 50% of the time when there is a file error.� Then by statistical inheritance which assumes the error messages are representative of their classes, the probability of "no network response" when the network is down is 0.9*0.2 = 0.18, and the probability of "error in file routines" when there are bugs in the system is 0.8*0.5 = 0.4, the probability of "no network response" when there are bugs in the system is 0.01*0.2 = 0.002, and the probability of "error in file routines" when the network is down is 0.01*0.5 = 0.005.

Then with a Naive Bayes approach, we can calculate the likelihoods of the three hypotheses "network is down", "bugs in the system", and "deception" after the download attempt as 0.10*0.18/0.02 = 0.9, 0.08*0.01/0.02 = 0.04, and 0.01*0.1/0.02 = 0.05.� We calculate likelihoods after the attempted modification of the executable as 0.9*0.005/0.06 = 0.075, 0.04*0.4/0.06 = 0.267, and 0.05*0.1/0.06 = 0.083.� So the contradictory clues have the effect of making deception more plausible than the other hypotheses even though its initial likelihood is low.

Connecting Generic Excuses with Deception Methods

These generic excuses differ in their compatibility with the defensive deception methods we enumerated in Table 1.� We can summarize compatibility in Table 2, where 0 means incompatible, 5 means indifferent, and 10 means compatible.� Compatibility includes both logical preconditions (network excuses only work for network actions) and statistical tendencies (hackers aren't usually concerned about redirecting your commands).� Again, these were obtained by expert-systems methodology and reflect the degree to which the deceptive act could be caused by the excuse.

Table 2: Compatibility of deception methods with generic excuses.

Hypothesis / Deception	Net down	Test	Bugs	Comms faulty	Policy change	Hacked	Joker	Deception
object	0	3	3	0	6	8	8	10
location-from	0	5	2	0	5	8	6	5
location-to	0	7	5	0	5	8	8	10
direction	0	0	3	0	0	8	5	5
frequency	8	10	8	8	4	1	5	1
time-at	0	8	2	2	0	8	2	5
time-through	0	8	5	8	5	8	8	8
cause	10	10	10	7	10	5	5	10
purpose	0	0	0	0	10	0	8	10
preconditions	0	7	10	0	8	0	10	10
ability	10	5	5	5	2	1	10	1
accompaniment	0	5	2	0	0	10	5	7
content	0	2	5	7	2	10	8	7
measure	0	5	2	0	5	2	3	7
value	3	8	5	10	2	10	8	7
effect	10	0	10	5	3	8	10	10

Putting Tables 1 and 2 together with the a priori probabilities of the excuses, we can express the suitability of a particular deception method i for achieving a particular generic excuse j as �where a is the a priori suitability of a given deception method, c is the compatibility of that deception method with a generic excuse, and r is the a priori likelihood of the excuse condition in a random computer system.� For the parameters in Tables 1 and 2, the best excuse-deception pairs are shown in Table 3.� It is reasonable� to use only those deception-excuse combinations exceeding a given threshold.

More than one generic excuse can be used, which is helpful since few excuses defend against all attacks; for instance, "The network is down" will not excuse the refusal to display a local directory.� However, using multiple excuses decreases significantly the trust of the attacker as suggested by our previous example.� The excuses enumerated above are mostly independent, so the simultaneous occurrence of more than one will have an a priori probability close to the product of their probabilities and likely be small, which helps the "deception" excuse which has weak but broad support.

Table 3: The best deception-excuse combinations as per our parameter estimates.

Deception method	Generic excuse	Suitability
effect	net down	10
cause	testing	9
cause	net down	9
effect	bugs	8
value	testing	7.2
cause	bugs	7.2
time-through	testing	6.4
frequency	testing	6.4
frequency	network down	6.4
value	comms. faulty	5.4

Suspiciousness of the Attacker

We assume an intrusion-detection system [24] can provide the data to calculate a probability that a user is an attacker.� This requires recognizing both misuse signatures (as in Snort) and anomalies, for both user-session information and all-users global information (in case a malicious user logs in under several names).� Misuse signatures generate high (but not certain) probabilities of attack, and anomalies such as high system load give conditional probabilities of attack that can be estimated from test data.� [22] implemented the architecture we propose to use in which both network-based and host-based open-source intrusion-detection systems wrote into log files read and used by a deception engine.

�

COUNTERPLANNING

With our estimate of how effective a set of deceptions can be, we can construct "counterplans" [7] to foil attackers.� This can be done in advance of any attack (if we can anticipate certain common attacks) or dynamically.� Good counterplanning tries to minimize damage to our computer systems.� [25] proposed a theory of piecemeal changes to thwart a detailed attack plan, but such ad hoc events are unlikely to be as convincing as the integrated deception provided by generic excuses.

The Penalty for Deceiving Non-Attackers

Our counterplans must specify what generic excuses we will use and when.� We should not apply an excuse until we are rather sure we are under attack (except for honeypots, which serve no legitimate purpose and can be deceptive all the time).� On the other hand, the later we apply the excuse, the more likely an attacker will still achieve their goals, since the closer people are to a goal, the harder they try to achieve it, and the more likely they will find some way to achieve it that we have not anticipated.� So some planning is necessary.

Let us assume a ratio r between the value of a legitimate user achieving their goals on this computer system to the value to the attacker of achieving their own goals; we expect this to be a small fraction.� Assume is the probability that the sequence of the first k actions by the user is malicious, obtained from the abovementioned intrusion-detection system.� Assume C is the cost to us of the attacker achieving their goals, and �the probability that they will achieve their goals.� Then the expected cost to us is

�.� This must be positive to justify trying to decrease the probability the attacker will achieve their goals, or that , so this gives an absolute criterion for deception early in an attack.

These formulas can be refined further.� Any intrusion-detection system we use to measure attacker suspiciousness will have "false positives" when a legitimate user accidentally creates a known attack signature or anomaly.� Honest people who have been tricked tend to react strongly, more so than to setbacks occurring without deceit.� So false positives could be ranked as extra costly in the formula.

A Metric for Counterplanning

As for the planning metric we wish to maximize, for most attacks this can be the probability that the attacker gives up and goes away (a "survival rate").� [25] proposed maximizing the amount of time wasted by the attacker, but this requires considerable knowledge of the attack plan, and could be dangerous: The more time an attacker spends on our system, the more likely they will recognize deceptions and figure countermoves (and perhaps tell other hackers).� So we assume we want to maximize the probability that an attacker gives up, a monotonic function of the probability of all the non-deception hypotheses.� All the generic excuses mentioned encourage an attacker to give up since they describe serious obstacles preventing their goals from being achieved.� In addition, effects can be intensified by additional deceptions from Table 1 like extra delays in system responses.� (Giving up on one system may encourage the attacker to attack another system of ours, but we should install similar deceptive methods on all our systems.)

Attackers (and legitimate users) can differ individually in their responses to adversity and deception (in patience, aggressiveness, etc.) [20].� It is helpful to identify through experiments those useful clusters of people having similar parameters to better adapt deceptions to personality.� "Personality tests" like the Myers-Briggs identify such standardized indicators; we expect that attackers will generally score high on the "thinking" (vs. "feeling") dimension and "sensing" (vs. "intuition") dimension, but will vary on the extraversion/introversion and judging/perceiving dimensions.� This can be exploited to predict their relative likelihood of giving up if we know something about them, say from observing how they respond to minor setbacks during their attack.� But we believe individual influences will be minor with the unsubtle generic excuses we have proposed.

So given this planning metric and some hypotheses about the attackers, we should use deception as soon as the cost formula says it is advantageous to do so, and then use it consistently from then on.� As to which generic excuses to use, this depends on the kinds of attack expected.� We can get statistics and simulate the expected effectiveness of our deceptions on the attack plans as [25] did.� This optimization problem can be addressed with many classic techniques in operations research and artificial intelligence such as A* search or genetic algorithms.� Several ideas can simplify the problem.� First, absolute optimality is not critical; we can use heuristics to "satisfice" and get a near-optimal solution.� Second, attackers have a limited set of goals they wish to achieve (the three most important are substituting their own copy of an operating system at a site, systematically searching for secrets, and swamping a resource with transactions so as to render it useless) and a limited set of methods they use to achieve those goals.� So it is reasonable to enumerate those plans and figure ways to interfere with each in advance rather than making decisions during an attack.� Third, we can learn from experience: We can see how volunteers or even attackers respond to methods we try, and use this to improve our modeling.

The attacker with a single goal like root compromise can be contrasted with the spy-like attacker that wants to reconnoiter our system and search for secrets, like those for whom our demonstration honeypot was designed.� Then the time-wasting metric is appropriate.� With information scattered randomly about, the rate at which the attacker acquires information will be roughly proportional to the amount of information content they have not yet examined, or dI/dt = K(1-I) where I is the fraction of the information content of the computer system that the attacker knows about, t is time, and K is a constant.� Solving this equation we get .� So we expect that a spy-like attacker will leave when the slope of this curve, decreases to the rate at which the spy values their expenditure of time.� We can lengthen their stay by increasing the time to obtain information or increasing the apparent amount of information on the system.� For instance for seven students assigned to explore our honeypot site, they visited 9, 11, 16, 25, 55, 67, and 95 pages respectively.� Assuming they value their time equally, the differences must be due to the constant representing the amount of value they obtain from the site.�� We can solve for the ratios of the constants for each pair of students.� For instance, between the most patient and least patient students, the ratio is approximately 5.9.�

Example: Deceptions against Rootkit Installation

Once a generic excuse has been chosen, we must decide when to use it.� For example, consider a rootkit-installation attack on a computer system.� An attacker can use many� methods (e.g. buffer overflows) to obtain administrator privileges.� But the basic plan remains the same: (1) Find a vulnerability, (2) escalate privileges, (3) copy their own software onto the target system, and (4) install it.� This entails subtasks like downloads by a file-transfer utility, decompression of the downloads, copy of the consequent files to specific locations in the operating system, and testing of the resultant installation.� We can itemize suitability of the steps of this plan for each category of generic excuses in Table 4.� Again, 0 means unsuitable and 10 means totally suitable.� These come from common sense: For instance, you cannot blame the network for local events, and you cannot blame communications for compiled operations.

Table 4: Compatibility of system commands with the generic excuses.

Generic excuse / Attacker action	Network down	Testing	Bugs	Comm. faulty	Policy change	Hacked	Joker	Deception
1. scan ports	10	5	5	10	10	5	8	10
2. connect at port	10	5	7	5	7	5	6	7
3. buffer overflow	0	10	10	10	7	5	2	10
4. file transfer from another site	10	5	5	5	10	7	8	7
5. decompress file	0	7	5	0	5	5	5	5
6. move file	0	7	5	2	8	5	5	7
7. test operating system	2	10	10	2	7	10	8	10

In deciding whether to use a deception-excuse pair, these suitabilities provide an additional multiplicative factor.� For instance, the excuse of "bugs" on the computer system should be used for the third and seventh actions where it is strongly appropriate.� It may also be appropriate at the fourth, fifth, and sixth actions, or at least one of them as a way to reinforce the message for the third action; partly suitable actions entail optional excuses.� As for at the second action, connecting at a port after a port scan is not a certain indicator of an attack (a legitimate user may just be logging in then), but a reasonable probability of maliciousness might be 0.2, and a reasonable value of �for many systems is 0.1.� So the inequality for justifying deception holds since 0.2+(0.1*0.2)-0.1 > 0, although its worth is decreased by the 0.7 suitability.

Attack analysis is better the more you know about attack plans; CMU's CERT publishes much useful information.� We can formulate those plans using a variety of planning models [10, 19]; [28] provides useful components for general such modeling.� Modeling also benefits if we estimate work necessary for an attacker to counter each of our deceptive ploys ("plan repair") [25].� For instance, if we lie and say a downloaded file is no longer present, the attacker will need to download it again, requiring a certain amount of wasted time.� Deceptive ploys requiring significant work to repair will significantly increase the probability that the attacker gives up because they increase the attacker's frustration.

Consistency

As discussed earlier, inconsistency decreases the probability of a generic excuse while leaving the probability of deception untouched.� That means that once an excuse is given, it should be given for all similar situations.� So for instance if we complain of "network problems" for a download, we should complain or give other similar evidence of network problems at every subsequent occasion that the same user requires the network.� We should also decrease the threshold for this deception for other users logged in at the same time since they may be persona of the same attacker.

Table 5 shows an example of consistency with the "you broke it" excuse.� FTP is a file-transfer utility.� IS is the intrinsic suspiciousness of the action, and CS is the cumulative suspiciousness taking context also into account.� Because user U2 tried to get something suspicious from site S earlier, the overly-soon second login for user U1 (likely to then be U2) also trying to get the same file is consistently denied file transfers, whereas earlier user U1 was not impeded.

Table 5: Example of reasoning for deception.

Time	User: Action	IS	CS	Response
800	U1: login	0.0	0.0	normal
805	U1: edit paper1	0.0	0.0	normal
808	?: scan ports	0.4	0.4	normal
812	U2: connect at port 6520	0.2	0.2	normal
813	U2: ftp encrypted file EF from site S	0.3	0.5	say file too big and network down
814	U1: ftp paper2 from site T	0.1	0.1	normal
815	U1: logout	0.0	0.0	normal
821	U1: login	0.1	0.1	normal
822	U1: ftp EF from S	0.3	0.7	say network down
823	U1: ftp encrypted file ER from� U	0.3	0.8	say network down

Similarly, we must also be consistent in the deception methods we choose.� If we lie that the network is down at some point, we should not later claim security policy is being violated by network transmissions, even though that is an equally valid excuse for disallowing network operations.� There are, however, exceptions to the principle of consistency.� If the file-transfer utility breaks only when the attacker tries to transfer suspicious files, the attacker will get suspicious.� It is better to use the "you broke it" strategy just once and prevent all file transfers after one suspicious one is attempted.� In general, it is best to avoid consistent unusual events.

It is true that an attacker not fazed by our first deceptive ploy is less likely to respond to a second or subsequent ploy.� Nonetheless, consistency is still important because attacker behavior in the face of surprises is often an attempt to confirm the surprise by a kind of� "hypothesis testing".� At that point, attacker reasoning no longer assumes that evidence probabilities are independent since a common explanation is being proffered.� For instance, the joint probability that both word-processing software and file copying don't work in a computer system is considerably more than the product of the individual probabilities because several common hypotheses explain both, including "file system not working" and "testing is being done".

How do we obtain good probabilities of what attackers will do with and without deception?� We need to do human-factors experiments, either with volunteers or with real attackers in honeypots.� We can foist deceptions on them or not and record statistics on how they respond.� We may also be able to infer the worth that attackers give to their own work time and to achieving their goals, something that will vary considerably between attackers and between situations.� Each time an attacker continues attacking our system, they are giving us an inequality between the expected value of a successful attack on our system and other targets; each time the attacker gives up an attack, we have an inequality in the other direction.� We can use methods of linear programming to find points on the boundary of the region of parameter space that satisfies these inequalities, and find its centroid.

CONCLUSIONS

Deception is a key part of human nature.� Cyber-attackers have used deceptions for a long time to achieve their goals.� It is only fair that defenders respond in kind.� So we need a good theory of how deception operates both offensively and defensively.� Simple defensive deceptions can be easy to implement, easier than configuring and implementing access controls on every resource of an information system.� We need to start deception planning for important information systems on a routine basis.� The techniques are not difficult, but do require some care to choose among the many options available.

ACKNOWLEDGMENTS

This work is part of the Homeland Security Leadership Development Program supported by the U.S. Department of Justice Office of Justice Programs and Office for Domestic Preparedness.�

REFERENCES

[1] Applegate, C., Elsaesser, C., & Sanborn, J., An architecture for adversarial planning, IEEE Transactions on Systems, Man, and Cybernetics, 20 (1), January/February 1990, pp. 186-194.

[2] Austin, J. L., How To Do Things With Words. (2nd ed., ed. by J.O. Urmson & M. Sbis). Oxford: Oxford University Press, 1975.

[3] Barber, R. S., and Kim. J., Belief revision process based on trust: agents evaluating reputation of information sources.� In Falcone, R., Singh, M., and Tan, Y.-H. (eds.), Trust in Cyber-Societies, LNAI 2246 (Berlin: Springer-Verlag, 2001), 73-82.

[4] Bell, J. B., & Whaley, B., Cheating.� New York: Transaction Publishing, 1991.

[5] Bishop-Clark, C., & Wheeler, D., The Myers-Briggs personality type and its relationship to computer programming.� Journal of Research on Technology in Education, Vol. 26, No. 3, pp. 358-370, Spring 1994.

[6] Bok, S., Lying: Moral Choice in Public and Private Life.� New York: Pantheon, 1978.

[7] Carbonell, J., Counterplanning: A strategy-based model of adversary planning in real-world situations, Artificial Intelligence, Vol. 16, 1981, pp. 295-329.

[8] Carley, K., Inhibiting adaptation, Proc. Command and Control Research and Technology Symposium, Monterey CA, USA, June 2002.

[9] Carofiglio, V., de Rosis, F., & Castelfranchi, C., Ascribing and weighting beliefs in deceptive information exchanges.� Proc. User Modeling, 2001, 222-224.

[10] Cohen, F., Simulating cyber attacks, defenses, and consequences.� http://all.net/journal/ntb/simulate/simulate .html, May 1999.

[11] Copeck, T., Delisle, S., & Szparkowicz, S., Parsing and case interpretation in TANKA. Conference on Computational Linguistics, Nantes, France, pp. 1008-1023, 1992.

[12] De Rosis, F., Castelfranchi, C., Carofiglio, V., & Grassano, R., Can computers deliberately deceive? A simulation tool and its application to Turing's imitation game.� Computational Intelligence, to appear 2004.

[13] Dunnigan, J. F., & Nofi, A. A., Victory and Deceit, second edition: Deception and Trickery in War.� San Jose, CA: Writers Club Press, 2001.

[14] Fillmore, C., The case for case.� In Universals in Linguistic Theory, ed. Bach & Harns, New York: Holt, Rinehart, & Winston, 1968.

[15] Ford, C. V., Lies! Lies!! Lies!!! The Psychology of Deceit.� Washington, DC: American Psychiatric Press, 1996.

[16] Fowler, C. A., & Nesbit, R. F., Tactical deception in air-land warfare.� Journal of Electronic Defense, Vol. 18, No. 6 (June 1995), pp. 37-44 & 76-79.

[17] Heuer, R. J., Cognitive factors in deception and counterdeception.� In Strategic Military Deception, ed. Daniel, D. C., & Herbig, K. L., New York: Pergamon, 1982, pp. 31-69.

[18] The Honeynet Project, Know Your Enemy. Boston: Addison-Wesley, 2002.

[19] Lowry, J., An initial foray into understanding adversary planning and courses of action.� Proc. DARPA Information Survivability Conference & Exposition II, Anaheim, CA, June 2001, vol. 1, pp. 123-133.

[20] Lydon, J., & Zanna, M., Commitment in the face of adversity: A value-affirmation approach, Journal of Personality & Social Psychology, Vol. 58, No. 6, June 1990, pp. 1040-1047.

[21] McClure, S., Scambray, J., & Kurtz, G., Hacking Exposed: Network Security Secrets and Solutions, third edition.� New York: McGraw-Hill Osborne Media, 2001.

[22] Monteiro, V., How intrusion detection can improve software decoy applications.� M.S. thesis, U.S. Naval Postgraduate School, March 2003, www.cs.nps.navy.mil/people/faculty/rowe/oldstudents/monteiro_thesis.htm.

[23] Nelms, H., Magic and showmanship: A handbook for conjurers.� Mineola, NY: Dover, 1969.

[24] Proctor, P. E., Practical intrusion detection handbook.� Upper Saddle River, NJ: Prentice-Hall PTR, 2001.

[25] Rowe, N. C., Counterplanning deceptions to foil cyber-attack plans.� Proc. 2003 IEEE Workshop on Information Assurance, West Point, NY, June 2003, 203-211.

[26] Smith, H. A., The compleat practical joker.� Garden City, NY: Doubleday, 1953.

[27] Sztompka, P., Trust.� London: Cambridge University Press, 1999.

[28] Templeton, S., & Levitt, K., A requires/provides model for computer attacks.� Proc. of the New Security Paradigms Workshop, Cork, Ireland, September 2000.

[29] Tognazzini, B., Principles, techniques, and ethics of stage magic and their application to human interface design.� Proc. Conference on Human Factors and Computing Systems (INTERCHI) 1993, Amsterdam, April 1993, pp. 355-362.

[30] Ward, D., & Hexmoor, H., Toward deception in agents.� Proc. 2^nd Intl. Joint Conf. on Autonomous Agents and Multi-Agent Systems, Melbourne, Australia, pp. 1154-1155, July 2003.

[31] Whaley, B., Towards a general theory of deception.� Journal of Strategic Studies, Vol. 5, No. 1, pp. 179-193, March 1982.

[32] Witten, I., & Frank, E, Data mining: Practical machine learning with Java implementations.� San Francisco, CA: Morgan Kaufmann, 2000.Top of Form

[33]