Top Banner
ACM NetGames 2008 Game Bot Identification Game Bot Identification based on Manifold Learning based on Manifold Learning KuanTa Chen Academia Sinica HsingKuo Pao NTUST HongChung Chang NTUST
36

Game Bot Identification Based on Manifold Learning

Jan 13, 2015

Download

Technology

In recent years, online gaming has become one of the most popular Internet activities, but cheating activity, such as the use of game bots, has increased as a consequence. Generally, the gaming community disapproves of the use of game bots, as bot users obtain unreasonable rewards without corresponding efforts. However, bots are hard to detect because they are designed to simulate human game playing behavior and they follow game rules exactly. Existing detection approaches either disrupt players’ gaming experiences, or they assume game bots are run as standalone clients or assigned a specific goal, such as aim bots in FPS games.

In this paper, we propose a manifold learning approach for detecting game bots. It is a general technique that can
be applied to any game in which avatars’ movement is controlled by the players directly. Through real-life data traces, we show that the trajectories of human players and those of game bots are very different. In addition, although game bots may endeavor to simulate players’ decisions, certain human behavior patterns are difficult to mimic because they are AI-hard. Taking Quake 2 as a case study, we evaluate our scheme’s performance based on real-life traces. The results show that the scheme can achieve a detection accuracy of 98% or higher on a trace of 700 seconds.
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Page 1: Game Bot Identification Based on Manifold Learning

ACM NetGames 2008

Game Bot Identification Game Bot Identification 

based on Manifold Learningbased on Manifold Learning

Kuan‐Ta Chen  Academia Sinica

Hsing‐Kuo Pao NTUST

Hong‐Chung Chang NTUST

Page 2: Game Bot Identification Based on Manifold Learning

2Kuan‐Ta Chen / Game Bot Identification based on Manifold Learning

Game Bots

Game bots: automated AI programs that can perform certain tasks in place of gamers

Popular in MMORPG and FPS games

MMORPGs (Role Playing Games)accumulate rewards in 24 hours a day 

break the balance of power and economies in game

FPS games (First‐Person Shooting Games)a) improve aiming accuracy onlyb) fully automated

achieve high ranking without proficient skills and efforts

Page 3: Game Bot Identification Based on Manifold Learning

3Kuan‐Ta Chen / Game Bot Identification based on Manifold Learning

Bot Detection

Detecting whether a character is controlled by a bot is 

difficult since a bot obeys the game rules perfectly

No general detection methods are available today

State of practice is identifying via human intelligence

Detect by “bots may show regular patterns or peculiar 

behavior”

Confirm by “bots cannot talk like humans”

Labor‐intensive and may annoy innocent players

Page 4: Game Bot Identification Based on Manifold Learning

4Kuan‐Ta Chen / Game Bot Identification based on Manifold Learning

Related Work

PreventionCAPTCHA(reverse Turing tests) [Golle et al; 2005]

DetectionProcess monitoring at client side [GameGuard]

Bot program’s signatures are keeping changing

Traffic analysis at the network [Chen et al; 2006]Remove bot traffic’s regularity by heavy‐tailed random delays

Aiming bot detection using DBN [Yeung et al; 2006]Specific to aiming bots that help aim the target accurately 

Page 5: Game Bot Identification Based on Manifold Learning

5Kuan‐Ta Chen / Game Bot Identification based on Manifold Learning

CAPTCHA in a Japanese Online Game

Page 6: Game Bot Identification Based on Manifold Learning

6Kuan‐Ta Chen / Game Bot Identification based on Manifold Learning

Our Goal of Bot Detection Solutions

Passive detection 

No intrusion in players’ gaming experience

No client software support is required

Generalizable schemes (for other games and other 

game genres)

Page 7: Game Bot Identification Based on Manifold Learning

7Kuan‐Ta Chen / Game Bot Identification based on Manifold Learning

Our Solution: Trajectory + Manifold Learning

Based on the avatar’s movement trajectory in game

Applicable for all genres of games where players control 

the avatar’s movement directly

Avatar’s trajectory is high‐dimensional (both in time 

and spatial domain) 

Use manifold learning to distinguish the trajectories 

of human players and game bots

Page 8: Game Bot Identification Based on Manifold Learning

8Kuan‐Ta Chen / Game Bot Identification based on Manifold Learning

The Rationale behind Our Scheme

The trajectory of the avatar controlled by a human 

player is hard to simulate for two reasons:

Complex context information: 

Players control the movement of avatars based on their 

knowledge, experience, intuition, and a great deal of 

environmental information in game. 

Human behavior is not always logical and optimal

How to model and simulate realistic movements (for 

game agents) is still an open question in the AI field. 

Page 9: Game Bot Identification Based on Manifold Learning

9Kuan‐Ta Chen / Game Bot Identification based on Manifold Learning

Bot Detection: A Decision Problem

Q: Whether a bot is controlling a game client giventhe movement trajectory of the avatar?

A: Yes / No?

Page 10: Game Bot Identification Based on Manifold Learning

10Kuan‐Ta Chen / Game Bot Identification based on Manifold Learning

Talk Progress

Overview

Data Description

Proposed Scheme

Pace vector construction

Dimension Reduction using Isomap

Classification

Performance Evaluation

Conclusion

Page 11: Game Bot Identification Based on Manifold Learning

11Kuan‐Ta Chen / Game Bot Identification based on Manifold Learning

Case Study: Quake 2

Choose Quake 2 as our case study

A classic FPS game

Many real‐life human traces are available on the Internet

more realistic than traces collected in experiments

Page 12: Game Bot Identification Based on Manifold Learning

12Kuan‐Ta Chen / Game Bot Identification based on Manifold Learning

A Screen Shot of Quake 2

Page 13: Game Bot Identification Based on Manifold Learning

13Kuan‐Ta Chen / Game Bot Identification based on Manifold Learning

Data Collection

Human traces downloaded from fan sites including GotFrag Quake, 

Planet Quake, Demo Squad, and Revilla Quake Site

Bot traces collected on our own Quake serverCR BOT 1.14

Eraser Bot 1.01

ICE Bot 1.0

Totally 143.8 hours of traces were 

collected

Page 14: Game Bot Identification Based on Manifold Learning

Aggregate View of Trails (Human & 3 Bots)

Human CR Bot

Eraser ICE Bot

Page 15: Game Bot Identification Based on Manifold Learning

15Kuan‐Ta Chen / Game Bot Identification based on Manifold Learning

Trails of Human Players

Page 16: Game Bot Identification Based on Manifold Learning

Trails of Eraser Bot

Page 17: Game Bot Identification Based on Manifold Learning

Trails of ICE Bot

Page 18: Game Bot Identification Based on Manifold Learning

18Kuan‐Ta Chen / Game Bot Identification based on Manifold Learning

Talk Progress

Overview

Data Description

Proposed Scheme

Pace vector construction

Dimension Reduction using Isomap

Classification

Performance Evaluation

Conclusion

Page 19: Game Bot Identification Based on Manifold Learning

19Kuan‐Ta Chen / Game Bot Identification based on Manifold Learning

The Complete Process: Overview

Decision

Step 1. Pace Vector Construction

Step 2. Dimension Reduction with Isomap

Step 3. Supervised classification

Page 20: Game Bot Identification Based on Manifold Learning

20Kuan‐Ta Chen / Game Bot Identification based on Manifold Learning

Step 1. Pace Vector Construction

For each trace sn , we compute the pace (distance) in 

successive two seconds by

We then compute the distribution (histogram) of paces 

with a fixed bin size by

where B is the number of bins in the distribution.

Fn = (fn,1, fn,2, . . . , fn,B)

ksn,i+1 − sn,ik =p(sn,i+1 − sn,i)T (sn,i+1 − sn,i)

Page 21: Game Bot Identification Based on Manifold Learning

21Kuan‐Ta Chen / Game Bot Identification based on Manifold Learning

Pace Vector: An Example

B is set to 200 (dimensions) in this work

Page 22: Game Bot Identification Based on Manifold Learning

22Kuan‐Ta Chen / Game Bot Identification based on Manifold Learning

Step 2. Dimension Reduction with Isomap

We adopt Isomap for nonlinear dimension reduction forBetter classifiaction accuracy

Lower computation overhead in classification

IsomapAssume data points lie on a manifold

1. Construct the neighborhood graph by kNN (k‐nearest neighbor)

2. Compute the shortest geodesic path for each pair of points

3. Reconstruct data by MDS (multidimensional scaling)

A mathematical space in which every point has a neighborhood which resembles Euclidean space, but in which the global structure may be more complicated. (Wikipedia)

Page 23: Game Bot Identification Based on Manifold Learning

23Kuan‐Ta Chen / Game Bot Identification based on Manifold Learning

A Graphic Representation of Isomap

Page 24: Game Bot Identification Based on Manifold Learning

24Kuan‐Ta Chen / Game Bot Identification based on Manifold Learning

PCA (Linear) vs. Isomap (Nonlinear)

Page 25: Game Bot Identification Based on Manifold Learning

25Kuan‐Ta Chen / Game Bot Identification based on Manifold Learning

Step 3. Classification

Apply a supervised classifier on the Isomap‐reduced 

pace vectors

SVM (Support Vector Machine) in our study

To decide whether a trajectory belongs to a game bot or 

a human player

Page 26: Game Bot Identification Based on Manifold Learning

26Kuan‐Ta Chen / Game Bot Identification based on Manifold Learning

Talk Progress

Overview

Data Description

Proposed Scheme

Pace vector construction

Dimension Reduction using Isomap

Classification

Performance Evaluation

Conclusion

Page 27: Game Bot Identification Based on Manifold Learning

27Kuan‐Ta Chen / Game Bot Identification based on Manifold Learning

Five Methods for Comparison

Method Data Input

kNN

Linear SVM

Nonlinear SVM

Isomap + kNN

Isomap + Nonlinear SVM

Isomap‐reduced Pace 

Vectors

Original 200‐dimension

Pace Vectors

Page 28: Game Bot Identification Based on Manifold Learning

Evaluation Results

Error Rate

False Postive Rate False Negative Rate

Page 29: Game Bot Identification Based on Manifold Learning

29Kuan‐Ta Chen / Game Bot Identification based on Manifold Learning

Addition of Gaussian Noise

Bot programmers can try to evade from detection by 

adding random noise into bots’ movement behavior

Evaluate the robustness of our schemem by adding 

Gaussian noise into bots’ trajectories

Page 30: Game Bot Identification Based on Manifold Learning

30Kuan‐Ta Chen / Game Bot Identification based on Manifold Learning

Evaluation Results

Error Rate

False Negative RateFalse Postive Rate

Page 31: Game Bot Identification Based on Manifold Learning

31Kuan‐Ta Chen / Game Bot Identification based on Manifold Learning

Cross‐Map Validation

Human movement may be restricted by the environment around 

him/her

Whether a classifier trained for a map can be used for detecting

bots on another map?

The Edge The Frag Pipe Warehouse

Page 32: Game Bot Identification Based on Manifold Learning

Evaluation Results

Error Rate

False Postive Rate False Negative Rate

Page 33: Game Bot Identification Based on Manifold Learning

33Kuan‐Ta Chen / Game Bot Identification based on Manifold Learning

Talk Progress

Overview

Data Description

Proposed Scheme

Pace vector construction

Dimension Reduction using Isomap

Classification

Performance Evaluation

Conclusion

Page 34: Game Bot Identification Based on Manifold Learning

34Kuan‐Ta Chen / Game Bot Identification based on Manifold Learning

Conclusion

We propose a trajectory‐based approach for detecting 

game bots.

The results show that the Isomap + nonlinear SVM 

approach performs good and stable results.

Human’s logic in controlling avatars is hard to simulate

we believe this approach has the potential to be a 

general yet robust bot detection methodology

Page 35: Game Bot Identification Based on Manifold Learning

35Kuan‐Ta Chen / Game Bot Identification based on Manifold Learning

Future Work

Include more spatial‐domain information in the pace 

vector

Validate our methodology on other games (game genres)

Page 36: Game Bot Identification Based on Manifold Learning

ACM NetGames 2008

Kuan‐Ta Chen

http://www.iis.sinica.edu.tw/~ktchen

Thank You!Thank You!