The Scientists Are Training Smart AI To Use Against Adversarial Attacks

Saanvi Araav - May 31, 2019


The Scientists Are Training Smart AI To Use Against Adversarial Attacks

There is this new type of hack with the name of adversarial attacks and scientists are training AI to deal with it

There is this new type of hack with the name of adversarial attacks. It is very dangerous and could input data in a maliciously way that will trigger glitch in our AI.

Adversarial Attacks Target The AI

In fact, these adversarial attacks could come in many different forms. Might be it will take on the form of audio input from a song that will take control of your Alexa, or it will become the roadside sign's pattern to confuse all those self-driving vehicles. At the lightest, they could be malicious, while at worst they could be very dangerous.

The-Scientists-Are-Training-Smart-Ai
The hackers aim at the vulnerabilities of the AI network

In the case of this type of attacks, those normal cyber defenses would not be able to prevent it at all. These systems won't be effective, because the hackers aim at the vulnerabilities of the AI network, not security loophole nor a human mistake. To deal with this dangerous possibility, a group of researchers from Carnegie Mellon University has been training the AI to prevent them from falling to these attacks. And in the process, they have also discovered the mechanic of those adversarial attacks as well.

A scientist - Zico Kolter has also said that certain AIs are just too smart and their intelligence is turning back at them. They have been developed in a way that they would be able to spot those pattern that human would miss, then interpret them as their instructions. That is the loophole that the hackers will exploit, even though the human developer thinks that these AIs are perfect.

The Training Data Set

Ai-Talent-Training-Lead
The scientists have made a special training data set.

To train the AIs against those hackers, scientists have made a special training data set. This set includes those images that to us is a certain thing, but then the computer thinks they are something else. For example, to us, that particular picture is a dog, but then the AI will think that it is a cat-like fur thing. The team has also labeled the images with the interpretation of the AI. After that training, the system was able to halve of their original time to identify the images.

From the experiment, scientists have noted that the recognition AI has two classification modes. One looks for the subtle microfeatures that human couldn't see, while the other looks for macro features like tail and ears etc. The adversarial attacks form on the basis of those microfeatures.

Tags

Comments

Sort by Newest | Popular

Next Story