Machine Learning in Game Development

Hi there, my name is Denis. I'm a software engineer at My Games. Today I'd like to share insights and discuss a prevalent issue in the gaming world – cheating. More importantly, I want to delve into how machine learning can be our ally in preserving the genuine essence of competition and fun in video games.

Trust is super important when playing games. Players are not just watching from the side. They are really part of the game. They put a lot of time, work and emotion into these virtual worlds. It is more than just playing. They try to understand how the game works. They think of ways to improve. And they also make friends inside the game.

On top of that, they sometimes spend money on things within the game to make their experience more enjoyable. However, when they come across players who cheat, it is more than just a small issue. It can make them feel like the game isn't being fair for everyone. It becomes really difficult for them to enjoy the game when they think that everyone is not following the same rules.

Keeping players interested for a long time is important for online games. In order to stay alive and do well. If players keep coming back, the game community stays alive and games can make more money. This is especially true for games where players buy things inside the game or play regularly to play. If there are cheaters, it can ruin the fun for other players. Instead of being fun or relaxing, games become annoying and players become frustrated.

There are various ways players can cheat in video games. Game developers need to identify whether a game might be vulnerable or have loopholes as well as identifying certain gameplay rules that could potentially be sidestepped. A prevalent hacking method in online games involves manipulating a participant state. Since a significant portion of the gameplay occurs on the player's own computer, hackers tamper with this by altering the game's local data. For instance, they might modify the bullet count in a gun's magazine, enabling the player to fire continuously without reloading, or amplify the available in-game items like grenades or health packs. There are more complex ways of cheating in games. Aimbot, aim assistant, or automatic aim — all these are different names for the same thing. Aimbot helps the user to track other targets by controlling the player's aiming.

The cheat operates by extracting game stage information that other players can't access. With this information, the bot then assists the cheater in aiming. By using an aimbot, a user surpasses human capabilities in shooter games.

Another cheat, called TRIGGERBOT, works the opposite way. While an aimbot automatically directs the crosshair at an opponent, TRIGGERBOT simply auto-fires when the player's crosshair is on an opponent. This means the player still has to aim at the enemy themselves, but as soon as they do, the TRIGGERBOT instantly activates, providing a perfect shot without delays that might arise due to human reflex.

Due to the peer-to-peer approach, each player's client has to contain up-to-date information about all other players. Naturally, the game standard interface hides all this information, showing only what the developers intended. Unfortunately, the mere presence of this information on the client paves the way for a type of cheating called overlay. One of these is wall hack, where a player can see the enemies through obstacles, be it walls, boxes or anything else. This gives the user an unfair advantage as they can prepare in advance for an encounter with an opponent, or even kill them while remaining unnoticed. Another similar example is radar hack. Instead of making walls transparent, this cheat displays the location of all players on the in-game radar or mini-map.

Of course there are other types of cheating as well, but let's move to the next point. Many of the often mentioned types of cheating can be controlled using traditional methods. We refer to this method as the rule-based model. The basis of this model is a set of rules that assess a player's actions for fairness. Let's delve deeper.

Players should send reports after each session. These reports include data about the actions of all participants. For example, we can see how many consumable items each player used by comparing this information with the player's account state on the server. We can check whether they had such resources. Another example is analyzing the number of shots and the time between them. This allows us to determine if the player adhered to the necessary reload time and whether the fight rate is consistent with the standards.

Some rules may directly indicate the use of cheats. Other rules provide only indirect evidence, for instance, temporary internet connection issues can affect data. So, it is a good idea to implement a cheating rating system, where each rule adds a certain number of cheating points. Some rules may contribute more points, while others add fewer, depending on the severity of the offense. If a player behaves fairly, their cheating rating decreases over time. Thus, occasional network issues might give a player a small number of points, which will be offset by fair gameplay.

A genuine cheater, however, will quickly accumulate enough points to raise suspicion. However, not all cheats can be detected through rules. For example, detecting the use of aim bots is challenging. If we merely analyze the ratio of headshots to the total number of shots, we risk falsely accusing a skilled player of cheating. This is where machine learning methods come into play.

Machine learning offers a powerful solution for complex cheating detection tasks, especially where traditional methods face limitations. One of the key advantages of machine learning is its ability to adapt to the dynamic and intricate behavior of players. Players continuously change their strategies and skills, which can make certain behavior patterns less apparent. Models based on machine learning are capable of adapting to these changes, identifying new forms of cheating.

Moreover, the volumes of data generated by games are often immense and machine learning can effectively process and analyze this information. This allows for the detection of subtle and intricate patterns in player's behavior, which might escape a basic analysis. With the right algorithm and sufficient data for model training, machine learning can provide high accuracy in detecting cheating, reducing the likelihood of false alarms.

It's also worth noting the proactive nature of machine learning. While many traditional methods respond to already known forms of cheating, machine learning can detect new, previously unknown methods. This adaptability and the ability to detect new threats quicker make it an ideal tool in the fight against cheaters.

Let me share my experience of using ML to detect cheaters. The idea behind my research was to use machine learning to identify anomalous behavior of players which would detect cheaters using aimbot. One of the first tasks was to figure out how to represent the behavior of players as digital data on which the model would be trained. Another task was to minimize the volume of this data to ease server workloads and reduce user traffic.

The challenge was to identify the minimum necessary data left to survive the game I'm working on is a specific mobile shooter. Players move through the scene as if on rails, hence the decision to not use movement data but focus more on aiming seemed evident. The theory was that difference in behavior between a player and an aimbot would be most noticeable in compensating for gun recoil. After numerous experiments, an optimal dataset was identified which produced commensurable results. A feature vector was determined which captured crosshair movement across both axes at 0.5 seconds before and at 0.25 seconds after the player fired as well as information on whether the shot hit an opponent. These feature vectors were only considered when a player was using a rapid-firing weapon.

I am primarily a game developer, not an experienced machine learning developer. To be honest, I faced numerous challenges during my research. The biggest problem was not having labeled data. To teach a model to sport if data is about cheating, you need data that is already labeled as cheating or not cheating. This means splitting the data into two groups, normal behavior and cheating behavior.

However, without a trained model, distinguishing between these two becomes a chicken and the egg problem. Okay, I eventually determined how to collect the data and represent it, but this data wasn't labeled. I need a technique to organize this data so it would be suitable for machine learning.

Sometimes after thinking about a problem for a long time, the answer is surprisingly simple. That's what happened there. I had some basic idea about the data I was looking at. I thought that the cheat data would be much smaller than the normal data, making it stand out. Thus, one can use anomaly detection on this initial data set to train a model for classifying data.

The pivotal aspect of utilizing machine learning to detect cheats was to formulate clear and relevant data depicting player behavior and selecting the most appropriate and simplest machine learning model for the task. However, for my vantage point, as a game developer, this indeed was the most challenging part of the entire process.

Despite the fact that the machine learning functionality for detecting cheaters, which I discussed today, is still in active development, its potential is already evident. In today's rapidly evolving technological world, machine learning has become a key tool for detecting and countering fraudulent activities in video games. Such advances foster optimism that the gaming industry will soon develop even more advanced and efficient anti-cheat systems.

Nevertheless, it's crucial to recognize that the advent of new methods in machine learning doesn't negate the importance of time-tested traditional approaches. Rule-based models, a classic method for detecting cheaters, remain relevant. These models granted in clearly defined rules complement and amplify the effectiveness of detection systems developed with machine learning algorithms. The blend of tried-and-true methods with innovations in machine learning can offer the best results in combating dishonest players.

I firmly believe that the future lies in flexible security systems where each component collaborates seamlessly with others, ensuring robust protection against cheating. And thank you for your attention today. I'm hopeful that as we continue to advance in machine learning technologies, we'll develop increasingly effective ways to maintain the integrity of the gaming experience we all hold dear. Let's keep pushing the boundaries to create more secure, enjoyable and fair games for everyone involved.