Adversarial examples are inputs to machine learning models that an attacker has intentionally designed to cause the model to make a mistake; they're like optical illusions for machines. In this post we'll show how adversarial examples work across different mediums, and will discuss why securing systems against them can be difficult. At OpenAI, we think adversarial examples are a good aspect of security to work on because they represent a concrete problem in AI safety that can be addressed in the short term, and because fixing them is difficult enough that it requires a serious research effort. To get an idea of what adversarial examples look like, consider this demonstration from Explaining and Harnessing Adversarial Examples: starting with an image of a panda, the attacker adds a small perturbation that has been calculated to make the image be recognized as a gibbon with high confidence. The approach is quite robust; recent research has shown adversarial examples can be printed out on standard paper then photographed with a standard smartphone, and still fool systems.

Israeli startup Binah.ai says it has developed technology that turns smartphones into health monitoring devices that can check vital signs including heartrate, oxygen saturation and respiratory rate. The new technology comes as medical care worldwide has been stretched thin by the pandemic and other, longer-term trends, spurring demand for telemedicine and cheaper, more convenient health monitoring solutions. The user just needs to look into the camera to let the company's system measure their vital signs. Our skin is constantly undergoing rapid changes in color, too subtle for us to notice, that reflect our body's physical state and functioning. "Basically we're following around the tiny color changes that are happening to the skin and the tiny color changes indicate the blood flow that is happening below the skin surface," Maman said.

"It's something that's a growing concern in the machine learning and AI community, especially because these algorithms are being used more and more," says Daniel Lowd, assistant professor of computer and information science at the University of Oregon. "If spam gets through or a few emails get blocked, it's not the end of the word. On the other hand, if you're relying on the vision system in a self-driving car to know where to go and not crash into anything, then the stakes are much higher." Whether or not a smart machine malfunctions, or is hacked, hinges on the very different way that machine learning algorithms'see' the world. In this way, to a machine, a panda could look like a gibbon, or a school bus could read as an ostrich.

The identical robotic explorers, Spirit and Opportunity, were able to trek up to 109 yards each Martian day. They found evidence for liquid water among many other things, with Opportunity traveling farther than any autonomous vehicle on any world: over 45 km (28 miles) over more than 5000 days. In 2004, NASA launched two exploration vehicles to the red planet: the Spirit and Opportunity rovers. These two Mars Exploration Rovers were originally designed for 90-day missions to image, explore, and investigate the Martian surface. Yet these twin solar-powered rovers far exceeded their design lifetimes.

Upon closer inspection, the satellite had noticed that an area that should have been shrouded in forest, was now barren. Within hours, a call had been made to a global conservation group, who mounted a legal case against the logging companies operating in the area. That process, historically, could have taken months of observing and recording changes. What's more, in remote areas such as the Ussuri Taiga in Russia's Far East, policing illegal logging operations have historically had little impact on the extraction of timber. But thanks to artificial intelligence (AI) and satellites, the ability to observe and respond to changes has become much faster.

Adversarial examples are inputs to machine learning models that an attacker has intentionally designed to cause the model to make a mistake; they're like optical illusions for machines. In this post we'll show how adversarial examples work across different mediums, and will discuss why securing systems against them can be difficult. At OpenAI, we think adversarial examples are a good aspect of security to work on because they represent a concrete problem in AI safety that can be addressed in the short term, and because fixing them is difficult enough that it requires a serious research effort. To get an idea of what adversarial examples look like, consider this demonstration from Explaining and Harnessing Adversarial Examples: starting with an image of a panda, the attacker adds a small perturbation that has been calculated to make the image be recognized as a gibbon with high confidence. An adversarial input, overlaid on a typical image, can cause a classifier to miscategorize a panda as a gibbon.

The car comes to a stop sign it's passed a hundred times before – but this time, it blows right through it. To you, the stop sign looks exactly the same as any other. But to the car, it looks like something entirely different. Minutes earlier, unbeknownst to either you or the machine, a scam artist stuck a small sticker onto the sign: unnoticeable to the human eye, inescapable to the technology. The tiny sticker smacked on the sign is enough for the car to "see" the stop sign as something completely different from a stop sign.

It could be the moment scientists accidentally put themselves out of a job. Physicists have revealed artificial intelligence software was used to run a complex experiment. The experiment, developed by physicists from ANU and UNSW ADFA, created an extremely cold gas trapped in a laser beam, known as a Bose-Einstein condensate, replicating the experiment that won the 2001 Nobel Prize. The experiment created an extremely cold gas trapped in a laser beam, known as a Bose-Einstein condensate, replicating the experiment that won the 2001 Nobel Prize. Bose-Einstein condensates are some of the coldest places in the Universe, far colder than outer space, typically less than a billionth of a degree above absolute zero.