CarderPlanet
Professional
- Messages
- 2,549
- Reaction score
- 724
- Points
- 113
Why doesn't game theory help you deal with threats?
Scientists from the Federal Polytechnic School of Lausanne in Switzerland have proposed a fundamentally new mechanism for protecting against cyber attacks and fraud for artificial intelligence systems. This method improves the reliability and predictability of neural networks and other machine learning algorithms.
The method is based on a modernized training scheme. Traditionally, the process is based on the principles of game theory, which, according to experts, have a number of drawbacks.
The new technology can ensure the security of almost all services that use the capabilities of AI: from video hosting sites like YouTube to unmanned vehicles.
The main problem with the standard approach is that the fight against intruders has always been considered a zero-sum game. If one side wins, the other side loses. This makes it much harder to train AI to defend itself.
YouTube, for example, contains a huge number of videos, which are completely impossible to analyze manually. AI solves this problem by classifying videos automatically and checking whether they meet certain standards.
However, classification systems are a rather vulnerable mechanism. Fraudsters, using tactics called "adversarial" in game theory, can introduce barely noticeable changes in the video. For example, add background noise that confuses the AI. This is how obscene videos are "checked" - even parental controls don't help.
Scientists decided to abandon the old paradigm in favor of a continuously adapting model. The AI defender and potential attacker will now perform different target functions, which will allow you to simulate realistic attack scenarios.
Thanks to this training, neural networks learn to counteract a variety of threats-from harmless fraud for fun to purposeful hacking of large systems.
The technology has already been successfully tested on image and video recognition mechanisms. In the future, they want to adapt it for other tasks as well.
Scientists from the Federal Polytechnic School of Lausanne in Switzerland have proposed a fundamentally new mechanism for protecting against cyber attacks and fraud for artificial intelligence systems. This method improves the reliability and predictability of neural networks and other machine learning algorithms.
The method is based on a modernized training scheme. Traditionally, the process is based on the principles of game theory, which, according to experts, have a number of drawbacks.
The new technology can ensure the security of almost all services that use the capabilities of AI: from video hosting sites like YouTube to unmanned vehicles.
The main problem with the standard approach is that the fight against intruders has always been considered a zero-sum game. If one side wins, the other side loses. This makes it much harder to train AI to defend itself.
YouTube, for example, contains a huge number of videos, which are completely impossible to analyze manually. AI solves this problem by classifying videos automatically and checking whether they meet certain standards.
However, classification systems are a rather vulnerable mechanism. Fraudsters, using tactics called "adversarial" in game theory, can introduce barely noticeable changes in the video. For example, add background noise that confuses the AI. This is how obscene videos are "checked" - even parental controls don't help.
Scientists decided to abandon the old paradigm in favor of a continuously adapting model. The AI defender and potential attacker will now perform different target functions, which will allow you to simulate realistic attack scenarios.
Thanks to this training, neural networks learn to counteract a variety of threats-from harmless fraud for fun to purposeful hacking of large systems.
The technology has already been successfully tested on image and video recognition mechanisms. In the future, they want to adapt it for other tasks as well.
