AI Players in This Simulated Sports Game Are Literally Psyching Each Other Out and It's Hilarious

The study explores what happens in the case of hackers' adversarial attacks.
Loukia Papadopoulos

Artificial intelligence (AI) can be trained to do a lot of things, especially learn how to adjust to its environment. But like all things in life, sometimes this training can go wrong.

One way it can get off course is through hackers' adversarial attacks. A new study is now revealing that AI can be fooled through these attacks into behaving in a way it shouldn’t. One such case is with the two players in this AI game.

This study is executed in a world of simulated sports, where both of the competing AIs train by playing against each other. The blue bot is the expert here, and the red one is supposed to learn from it.

However, something unexpected happens as the red bot chooses to fall down and not play the game as it should. The system has now been hacked.

The blue bot ends up being confused and plays terribly. It wobbles around in a weird manner, and loses many games, even though it is supposed to be better.

This is a case of an adversarial attack. The system has been hacked and has been put off its right track. Studying these circumstances is important in keeping us safe from hackers.

Add Interesting Engineering to your Google News feed.
Add Interesting Engineering to your Google News feed.
message circleSHOW COMMENT (1)chevron
Job Board