A presentation at Black Hat USA in in Las Vegas, NV, USA by Kelly Shortridge
BIG GAME THEORY HUNTING THE PECULIARITIES OF HUMAN BEHAVIOR IN THE INFOSEC GAME Kelly Shortridge (@swagitda_) Black Hat 2017
I’m Kelly
This is game theory
It’s time for hunting some game theory 4
Do you believe bug-free software is a reasonable assumption? 5
Do you believe wetware is more complex than software? 6
Traditional Game Theory relies on the assumption of bug-free wetware 7
Behavioral Game Theory assumes there’s no such thing as bug-free 8
“Think how hard physics would be if particles could think” —Murray Gell-Mann 9
“Amateurs study cryptography, professionals study economics” —Dan Geer quoting Allan Schiffman 10
This is what you’ll learn: 11
Let’s go hunting to find out why 13
I. What is Game Theory?
tl;dr – game theory is a mathematical language used to describe scenarios of conflict and cooperation 15
Game theory is more about language than theory Use it as a engendering tool, not as something to dictate optimal strategies 16
GT applies whenever actions of players are interdependent Strategic scenarios include many types of games, with different “solutions” for each 17
Zero Sum Games Lou Levit 18
Non-Zero Sum Games 19
Negative Sum Games 20
Positive Sum Games 21
Complete vs. Incomplete Information 22
Perfect vs. Imperfect Information Paul Green 23
Information Symmetry vs. Asymmetry Dakota Corbin 24
Defender Attacker Defender Games
Sequential games in which sets of players are attackers and defenders Assumes people are risk-neutral & attackers want to be maximally harmful 26
First move = defenders choosing a defensive investment plan Second move = attackers observe the defensive preparations & choose an attack plan 27
Nash Equilibrium is often used to “solve” games. This is bad. 28
Nash Equilibrium = optimal outcome of a noncooperative game Players are making the best decisions for themselves while taking their opponent’s decisions into account 29
Prisoner’s Dilemma Player 1 Player 2 Confess Refuse Confess -2, -2 0, -4 Refuse -4, 0 -1, -1 30
Nash Equilibrium is based on a priori reasoning Assumes rational, all-knowing players Assumes others’ decisions don’t affect you 31
People have applied Nash Equilibrium to infosec over the years… 32
Defender should play extremely fast so the attacker drops out of the game Better to invest in security than not invest, regardless of attacker strategy (wow!) Just apply tons of mathematical equations! 33
II. New defensive framework
Use GT for its expressive power in describing a framework for the infosec game Look at data outside GT, e.g. from experiments in domains similar to infosec, to select correct assumptions 35
What game is infosec? Ruben Bagues
DAD game (continuous defense & attack) Non-zero-sum Incomplete, imperfect, asymmetrical information Sequential / dynamic 37
This is a (uniquely?) tricky game 38
Have you heard infosec described as a “cat and mouse” game before? 39
Traditional Game Theory doesn’t allow for those… …or most of the characteristics of the “infosec game” 40
Assumes people are rational (they aren’t) Assumes static vs. dynamic environments Can’t ever be “one step ahead” of your adversary Deviations from Nash Equilibrium are common 41
“I feel, personally, that the study of experimental games is the proper route of travel for finding ‘the ultimate truth’ in relation to games as played by human players” —John Nash 42
Behavior-based framework Dayne Topkin
Experimental – how do people actually behave? People predict their opponent’s moves by either “thinking” or “learning” 44
Thinking = modeling how opponents are likely to respond 45
Our brains work like volatile memory 46
Working memory is a hard constraint for human thinking Enumerating steps past the next round is hard Humans kinda suck at recursion 47
Learning = predicting how players will act based on prior games / rounds 48
Humans learn through “error-reinforcement learning” (trial & error) People have “learning rates,” how much experiences factor into decision making Dopamine neurons encode errors 49
Veksler & Buchler study 200 consecutive “security games” across 4 strategies Different learning rates for attackers Tested # of prevented attacks for each strategy 50
Fixed strategy = prevent 10% - 25% of attacks Game Theory strategy = prevent 50% of attacks Random strategy = prevent 49.6% of attacks Cognitive Modeling strategy = prevent between 61% - 77% 51
Don’t be replaced by a random SecurityStrategyTM algorithm 52
III. Implementation 53
SWOT Analysis Scott Webb
101: Traditional SWOT Strengths, Weaknesses, Opportunities, Threats 56
Model SWOT for yourself in relation to your adversary Model SWOT for your adversary in relation to you 57
“The primary insight of GT is the importance of focusing on others – of putting yourself in the shoes of other players and trying to play out all the reactions…as far ahead as possible” – Adam Brandenburger 58
Strengths Weaknesses ▪ Understanding of target environment ▪ Inadequate budget ▪ Motivation to not be breached ▪ Limited employee training ▪ Lack of personnel Opportunities Threats ▪ Leverage new tech to allow for tear up/down ▪ Attackers can use new tech for scalability ▪ Increased board attention to get budget ▪ Hard to keep up with pace of new attack surface 59
201: Perceptual SWOT 60
For you and your adversary, consider: How can the strengths be weaknesses? How can the weaknesses be strengths? 61
Self vs. Other Reality Strength Perception Threat Opportunity 62 Strength Weakness Weakness Strength Weakness Perception Strength Reality Weakness Opportunity Threat
“Core rigidities” = deeply embedded knowledge sets that create problems Compliance, fixed security guidelines Top management can be the wrong people for an evolving environment 63
Attacker strength = having time to craft an attack Leverage that “strength” with strategies leading down rabbit holes and wasting their time 64
Attacker strength = access to known vulns Confuse them with fake architecture so they can’t be certain what systems you’re running 65
Thinking Exploitation
Thinking strategy: belief prompting Increase players thinking by one step 67
“Prompt” the player to consider who their opponents are & how their opponents will react Model assumptions around capital, time, tools, risk aversion 68
Your goal is to ask, “if I do X, how will that change my opponent’s strategy?” 69
A generic belief prompting guide: 70
How would attackers pre-emptively bypass the defensive move? What will the opponent do next in response? Costs of the opponent’s offensive move? Probability the opponent will conduct the move? 71
Example: A skiddie lands on one of our servers, what do they do next? 72
Perform local recon, escalate to whatever privs they can get Counter: priv separation, don’t hardcode creds Leads to: attacker must exploit server, risk = server crashes 73
Decision Tree Modelling Jessica Furtney
Model decision trees both for offense & defense Theorize probabilities of each branch’s outcome Creates tangible metrics to deter selfjustification 75
“Attackers will take the least cost path through an attack graph from their start node to their goal node” – Dino Dai Zovi, “Attacker Math” 76
25% 65% Reality Criminal Group Skiddies / Random 60% Priv Separation 0% Nation State 100% 10% 50% 50% Known exploit 30% Role sep 0% 60% Use DB on box 0% Win 40% Absorb into 25% botnet Tokenization, 70% segmentation 40% Anomaly detection 0% Win 77 98% 50% seccomp 10% 1day GRSec 5% 2% Elite 0day
Whiteboards + camera snaps (or “DO NOT ERASE!!!!”) Draw.io, Gliffy (plugs into Confluence) Google Docs (> insert drawing) PowerPoint (what I used) Visio (last resort) 79
Decision trees help create a feedback loop to refine strategy 80
Decision trees help for auditing after an incident & easy updating Serves as a historical record to refine decisionmaking process Mitigates “doubling down” effect by showing where strategy failed 81
Defender’s advantage = knowing the home turf Visualize the hardest path for attackers – how can you force them onto to that path? Commonalities on trees = which strategies mitigate the most risk across various attacks 82
Make decision trees the new “nice report” #WOCinTech
A new request for your pen-testers / red-team The ask: outline which paths they did or didn’t take, and why (a decision tree w/ explanations) Helps you see the attacker perspective of your defenses & where to improve 84
Learning Exploitation
Information asymmetry exploitation – disrupt attacker learning process Learning rate exploitation – introduce unreliability and pre-empt moves 86
Exploit the fact that you understand the local environment better than attackers 87
Falsifying Data Braydon Anderson
Defenders have info adversaries need to intercept Dynamic envs = frequently in learning phase Hide or falsify data on the legitimate system side 89
Macron Case Study Soroush Karimi
Allegedly used phishing tarpitting Signed onto phishing pages & planted bogus creds and info Obvious fakes in dumped documents 91
#wastehistime2016…but for hackers 92
Goal is to remove the attacker’s scientific method so they can’t test hypotheses (Pretend like hashtags are a thing and tweet #wastehackertime2017 with your own ideas) 93
Create custom email rejection messages Create honeydoc on the “Avallach Policy” Have response to suspicious emails be, “This violates the Avallach policy” Track when the doc is accessed 94
General strategy: create honeytokens that look to describe legitimate policies or technologies that would be useful in attacker recon 95
Non-Determinism Candice Seplow
Different behaviors at different times Can’t expect same result every time 97
ASLR is a non-deterministic feature, but highly deterministic in that it always works the same I want to amplify and extend it to higher levels 98
Raise costs at the very first step of the attack: recon Make the attacker uncertain of your defensive profile and environment 99
Attackers now design malware to be VM-aware #WOCinTech 100
Good: Make everything look like a malware analyst’s sandbox Better: Make everything look like a different malware analyst’s sandbox each time 101
Put wolfskins on the sheep
Mix & match hollow but sketchylooking artifacts on normal, physical systems 103
RocProtect-v1 – https://github.com/fr0gger/RocProtect-V1 Emulates virtual artifacts onto physical machine (see Unprotect Project as well) 104
VMwareServices.exe VBoxService.exe Vmwaretray.exe VMSrvc.exe vboxtray.exe ollydbg.exe wireshark.exe fiddler.exe \.\pipe\cuckoo cuckoomon.dll dbghelp.dll Mac addresses: “00:0C:29”, “00:1C:14”, “00:50:56”, “00:05:69” 105
system32\drivers\VBoxGuest.sys system32\drivers\VBoxMouse.sys HKLM\SOFTWARE\Oracle\VirtualBox Guest Additions C:\cuckoo, C:\IDA Program Files\Vmware 106
Make the IsDebuggerPresent function call always return non-zero Create fake versions of driver objects like .\NTICE and .\SyserDbgMsg Set KdDebuggerEnabled to 0x03 107
Load DLLs from AV engines using a Windows loader with a forwarder DLL ex64.sys (Symantec) McAVSCV.DLL (McAfee) SAUConfigDLL.dll (Sophos) cbk7.sys (Carbon Black) cymemdef.dll (Cylance) CSAgent.sys (Crowdstrike) 108
Deploy lightest weight hypervisor possible for added “wolfskin” https://github.com/asamy/ksm https://github.com/ionescu007/SimpleVisor https://github.com/Bareflank/hypervisor 109
Minimax Mike Wilson 110
Minimax / maximin = minimize the possible loss for a worst case maximum loss scenario 111
Want to find the minimum of the sum of the expected cost of protection and expected cost of non-protection 112
Don’t have a monoculture – diversity is strongly beneficial for protection Stochastic decisions may be better than deterministic From The Imitation Game: should only act on Enigma info some of the time, not all 113
Looking Ahead
Fluctuating infrastructure using emerging tech in “Infrastructure 3.0” Netflix’s Chaos Monkey https://github.com/Netflix/SimianArmy/wiki/Cha os-Monkey 115
Modelling attacker cognition via model tracing Prerequisite: how to begin observing attacker cognition 116
Preferences change based on experience Models incorporate the “post-decision-state” Higher the attacker’s learning rate, easier to predict their decisions 117
ΔUA = α (R – UA), where: ▪ UA = expected utility of an offensive action ▪ α = learning rate ▪ R = feedback (success / failure) 118
If α = 0.2, R = 1 for win & -1 for loss, then: ▪ ΔUA = 0.2(1- 0) = 0.2 Attacker is 20% more likely to do this again From here, you can adjust the learning rate based on data you see 119
Track utility values for each attacker action For detected / blocked actions, attacker action & outcome are known variables (so utility is calculable) Highest “U” = action attacker will pursue 120
IV. Conclusion 121
It is no longer time for some Game Theory 122
In fact, we’ve learned that GT is a language, not even a theory 123
Start with a SWOT analysis to gain perspective 124
Use thinking exploitation to improve threat modelling 125
Use learning exploitation to beleaguer your adversaries 126
Let’s work together to build strategies based on this behavioral framework 127
Next step – how to begin modeltracing attackers After that – predict attacker behavior 128
Try these at home – make your blue team empirical Worst case, random strategies beat fixed ones & are just as good as GT 129
“Good enough is good enough. Good enough always beats perfect.” —Dan Geer 130
Suggested reading ▪ David Laibson’s Behavioral Game Theory lectures @ Harvard ▪ “Game Theory: A Language of Competition and Cooperation,” Adam Brandenburger ▪ “Advances in Understanding Strategic Behavior,” Camerer, Ho, Chong ▪ “Know Your Enemy: Applying Cognitive Modeling in the Security Domain,” Veksler, Buchler ▪ “Know Your Adversary: Insights for a Better Adversarial Behavioral Model,” Abbasi, et al. ▪ “Deterrence and Risk Preferences in Sequential Attacker–Defender Games with Continuous Efforts,” Payappalli, Zhuang, Jose ▪ “Improving Learning and Adaptation in Security Games by Exploiting Information Asymmetry,” He, Dai, Ning ▪ “Behavioral theories and the neurophysiology of reward,” Schultz ▪ “Evolutionary Security,” and “Measuring Security,” Dan Geer 131
@swagitda_ /in/kellyshortridge kelly@greywire.net 132
We all groan when we hear it’s “time for some game theory,” but traditional game theory – modelling conflict and cooperation between rational decision-makers – still pervades how we think of defensive strategy as an industry. This primitive analysis is a disservice to defenders, who are facing humans (and who are, in fact, humans themselves), but are modelling their own actions and opponent’s actions based on the assumption of machine-like behavior.
In this session, I will examine traditional game theory and propose why behavioral game theory should take its place in the philosophy of defense. Next, I’ll review the first principles of game theory, through the lens of behavioral game theory, which empirically measures how humans actually behave in games, rather than assumes they will behave coldly rational.
I’ll explain the “rules” of the information security game and how traditional game theory is poorly suited to those conditions, along with the various behavioral models and why they are a superior fit. I’ll then explore the two primarily methods that play into how humans make decisions in games – “thinking” and “learning” and what empirical data from behavioral game theory studies suggests on how to improve thinking and learning, extrapolating to applications for infosec defenders.
Finally, I’ll present new insights from my own research, examining how defenders and attackers play the infosec game specifically, and bridging from theory to practice, to see how the lessons from behavioral game theory can be tangibly incorporated into defenders’ strategic decision making processes. I’ll conclude the session by outlining the practical steps for improving threat modelling, countering offensive moves, and deciding which products to use, so that defenders can start gaining the high ground in the infosec game.
Also presented at BSides PDX 2017.
PDF: https://swagitda.com/speaking/us-17-Shortridge-Big-Game-Theory-Hunting.pdf
The following resources were mentioned during the presentation or are useful additional information.
Citation of this talk by the BBC
Brandenburger, A. M., & Nalebuff, B. J. (2011). Co-opetition. Crown Business.
Huck, S. (Ed.). (2004). Advances in understanding strategic behaviour: Game theory, experiments and bounded rationality. Springer.
Veksler, V. D., & Buchler, N. (2016). Know Your Enemy: Applying Cognitive Modeling in Security Domain. In CogSci.
Abbasi, Y., Kar, D., Sintov, N., Tambe, M., Ben-Asher, N., Morrison, D., & Gonzalez, C. (2016, August). Know Your Adversary: Insights for a Better Adversarial Behavioral Model. In CogSci.
Payyappalli, V. M., Zhuang, J., & Jose, V. R. R. (2017). Deterrence and risk preferences in sequential attacker–defender games with continuous efforts. Risk Analysis, 37(11), 2229-2245.
He, X., Dai, H., & Ning, P. (2015, April). Improving learning and adaptation in security games by exploiting information asymmetry. In 2015 IEEE Conference on Computer Communications (INFOCOM) (pp. 1787-1795). IEEE.
Schultz, W. (2006). Behavioral theories and the neurophysiology of reward. Annu. Rev. Psychol., 57, 87-115.
Measuring Security by Dan Geer
Here’s what was said about this presentation on social media.
for free. You
can too.