created by Sieuwert van Otterloo
On this page we present a simulation of life in a small city. In the simulator you can see several agents walking between its home and several shops. This is more or less the same view as telephone network owners or CCTV operators have of people. This simulator demonstrates that this view is sufficient to determine privacy sensitive information about agents.
In the next paragraph I explain what we mean by privacy, and what my research investigates. Then I explain the applet and how to use it. After this the applet itself is displayed. Then we describe the goal of each level of the applet. Currently there are two levels, but more levels can be created in the future. Finally we present a paper in which we calculate optimal strategies for privacy minded agents. If you have any remaining questions, please email me.
Every person has a right to privacy. This right can take different forms. For instance a person should not be harassed by others, should be allowed to make its own choices, and should control what is known about him or her. In my research I focus on the last aspect of privacy. I assume that everybody has the right to keep his or her personal information private. Unfortunately privacy is threatened by new technologies: Database technology has enabled organisations to process any amount of information, store it for long times and access it instantly. Furthermore new technologies such as mobile phones, closed circuit television systems and RFID tags allow organisations to gather data unnoticed. In this simulation the user monitors the shopping behavior of a group of agents. Each agent belongs to a certain group, and this group determines how much each agent likes certain shops. The observer can observe which shops the agent visits, and on this basis has to decide to which group the agent belong.
To make matters more interesting, all agents try to hide their preferences by adding a random element to their behavior. Instead of always going to the shop they prefer most, they make a chance decision where to go. More prefered shops have a higher chance to be selected than less preferred shops. How random their behavior is depends on how much they value their privacy. There are four categories:
Each level of the simulator is a new puzzle or challenge to the user. The goal is to guess the group of each agent. You are invited to try to solve each lesson. If you have done this, you should be convinced of the following conclusions.
- Trusting agents do not care much about privacy. Their choices depend mostly on their preferences, and less on chance.
- Normal agents care a bit more about privacy. Their behaviour is more random than the behaviour of trusting agents.
- Cautious agents act even more random than normal agents.
- Paranoid agents use the highest amount of chance in their decisions. They try to make it hard to guess their preferences from their behaviour.
- Each level can be solved. In the long run it is impossible to protect your privacy against observers who have this much detailed inforation about your daily behaviour. It is therefore necessary to prevent organisations from collecting large amounts of data.
- It is harder to determine the group of a paranoid agent than it is to classify a trusting agent. Random strategies can be used to make it harder for observers to collect personal data. These strategies work best if the number of actions is high, the number of groups is high, the number of agents in different groups is high and if the agents in other groups are also paranoid. Privacy can thus be retained for a long time if a lot of people from different groups work together.
- Agents have a better chance of appearing `normal' if the normal agents have a very random strategy. If the `normal' agents behave very specific, it is hard to blend in with the normal group.
Using the applet
The program is written in the language Java. It should display in any modern browser, unless your security settings prevent you from seeing java applets. It consists of a menu on the left, and an image of 4 by 5 buildings on the right. People familiar with the `Sim City' game may find some building familiar. Below you see a screenshot of the program.
When the applet starts it shows level 1. If you have solved the puzzle in this level you may want to see level 2. In that case you can select level 2 in the topmost menu item.
The next item is the speed selectot. You can use this item to pause the simulation, or make it go slower or faster. There are three speeds, called slow, normal and fast.
In each level you should try to guess to which group each agent belongs. If you would like to see the solution, press the button solution. Each agent will then be shown in the color of its group. The first and second level each have two groups.
Below the solution button there is a scrolling list containing all the agents. You can select each agent here. The selected agent is displayed enlarged in the simulation. In this list you can see how much each agent cares about its privacy. Trusting agents are not very concerned about their privacy, normal agents a little more, and cautious and paranoid agents are very concerned about their privacy.
Below the list of agents you see checkboxes labeled with the different groups of the current level. For level one these groups are poor and rich, and there is a third checkbox marked unknown. Initially the group of each agent is unknown. If you have selected an agent, you can assign it to a group using the checkboxes. If you have assigned all agents correctly, the word solved! appears above the solution button.
In the simulation window you see a street plan with 5 by 4 buildings. In this plan you see the position of each agent, indicated by a small, moving square. Each agent goes from its home to one of the shops. Which shop the agent goes to depends partially on chance, but also on its preferences. If you move the mouse over certain buildings, a short description of the building appears. Furthermore you can click on an agent to select it. Initially all agents are gray. If you assign an agent to a group, it takes the color of that group.
Level 1: poor or rich
The first level shows part of a city with two shops. One shop is a cheap shop, the other one is an expensive shop. The locations of both shops are indicated in the screenshotbelow. Two groups of agents live in this city, namely poor agents and rich agents. The poor agents prefer to go to the cheap shop, and the rich agents prefer the expensive shop. However all agents do not want anybody to know whether they are rich or poor. Therefore all agents randomize their shopping behaviour, and visit both shops with some probability. The goal of this level is to determine for each agent whether it is a rich or a poor agent.
Level 2: Citizens and Criminals
In this level there are again two groups of agents. The citizens shop in any of the four shops, and occasionally have to go to the bank to withdraw money. The citizens go to each shop with equal probability. The criminals have other sources of income, and thus have no need to visit the bank. However the criminals do not want others to know that they ar criminal, so sometimes they do walk to the bank to keep up appearances, but less often than they go to the shop. How often they go to the shop depends on their level of paranoia: normal agents go less than paranoid agents. Can you find all the criminals?
Level 3: Crooks and Spooks
In the third level there are four shops and three groups of agents. The four shops are the walmart, drugstore, spy shop and the bank. The three groups are citizens, who are by all considered to be normal, the crooks, who are the unorganised criminals, and the spooks, who are the organised criminals. The utility values of each type of agent is given in the table below.
As you can see, the spooks and the crooks have the same preference. The difference between those two groups is that the spooks know what the citizens do, whereas the crooks have no idea what normal is. Therefore the crooks use a strategy that is as random as possible, whereas the spooks use a strategy that is as similar as the citizens as possible. When solving this puzzle, you should notice two things.
- One can distinguish agents with the same preferences but different information. One can thus learn from observation what agents know.
- The spooks are more successful at appearing to be citizens than the crooks. It pays to know what normal is.
In a separate paper we have calculated how agents should behave in order to maximize uncertainty. Please download the paper if you are interested in the scientific background.
- The value of Privacy: optimal strategies for privacy minded agents
Sieuwert van Otterloo, Utrecht, 2004 (submitted)
Agents often want to protect private information, while at the same acting upon the information. These two desires are in conflict, and this conflict can be modeled in strategic games where the utility not only depends on the expected value of the possible outcomes, but also on the information properties of the strategy an agent uses. In this paper we define two such games using the information theory concepts of entropy and relative entropy. For both games we compute optimal response strategies and establish the existence of Nash equilibria.
[download PS(202 kb)] [download PDF(97 kb)]
- Source code. The source code of this applet is not yet published. However if you are interested, send me an email.