Tag Archives: host
#431238 AI Is Easy to Fool—Why That Needs to ...
Con artistry is one of the world’s oldest and most innovative professions, and it may soon have a new target. Research suggests artificial intelligence may be uniquely susceptible to tricksters, and as its influence in the modern world grows, attacks against it are likely to become more common.
The root of the problem lies in the fact that artificial intelligence algorithms learn about the world in very different ways than people do, and so slight tweaks to the data fed into these algorithms can throw them off completely while remaining imperceptible to humans.
Much of the research into this area has been conducted on image recognition systems, in particular those relying on deep learning neural networks. These systems are trained by showing them thousands of examples of images of a particular object until they can extract common features that allow them to accurately spot the object in new images.
But the features they extract are not necessarily the same high-level features a human would be looking for, like the word STOP on a sign or a tail on a dog. These systems analyze images at the individual pixel level to detect patterns shared between examples. These patterns can be obscure combinations of pixel values, in small pockets or spread across the image, that would be impossible to discern for a human, but highly accurate at predicting a particular object.
“An attacker can trick the object recognition algorithm into seeing something that isn’t there, without these alterations being obvious to a human.”
What this means is that by identifying these patterns and overlaying them over a different image, an attacker can trick the object recognition algorithm into seeing something that isn’t there, without these alterations being obvious to a human. This kind of manipulation is known as an “adversarial attack.”
Early attempts to trick image recognition systems this way required access to the algorithm’s inner workings to decipher these patterns. But in 2016 researchers demonstrated a “black box” attack that enabled them to trick such a system without knowing its inner workings.
By feeding the system doctored images and seeing how it classified them, they were able to work out what it was focusing on and therefore generate images they knew would fool it. Importantly, the doctored images were not obviously different to human eyes.
These approaches were tested by feeding doctored image data directly into the algorithm, but more recently, similar approaches have been applied in the real world. Last year it was shown that printouts of doctored images that were then photographed on a smartphone successfully tricked an image classification system.
Another group showed that wearing specially designed, psychedelically-colored spectacles could trick a facial recognition system into thinking people were celebrities. In August scientists showed that adding stickers to stop signs in particular configurations could cause a neural net designed to spot them to misclassify the signs.
These last two examples highlight some of the potential nefarious applications for this technology. Getting a self-driving car to miss a stop sign could cause an accident, either for insurance fraud or to do someone harm. If facial recognition becomes increasingly popular for biometric security applications, being able to pose as someone else could be very useful to a con artist.
Unsurprisingly, there are already efforts to counteract the threat of adversarial attacks. In particular, it has been shown that deep neural networks can be trained to detect adversarial images. One study from the Bosch Center for AI demonstrated such a detector, an adversarial attack that fools the detector, and a training regime for the detector that nullifies the attack, hinting at the kind of arms race we are likely to see in the future.
While image recognition systems provide an easy-to-visualize demonstration, they’re not the only machine learning systems at risk. The techniques used to perturb pixel data can be applied to other kinds of data too.
“Bypassing cybersecurity defenses is one of the more worrying and probable near-term applications for this approach.”
Chinese researchers showed that adding specific words to a sentence or misspelling a word can completely throw off machine learning systems designed to analyze what a passage of text is about. Another group demonstrated that garbled sounds played over speakers could make a smartphone running the Google Now voice command system visit a particular web address, which could be used to download malware.
This last example points toward one of the more worrying and probable near-term applications for this approach: bypassing cybersecurity defenses. The industry is increasingly using machine learning and data analytics to identify malware and detect intrusions, but these systems are also highly susceptible to trickery.
At this summer’s DEF CON hacking convention, a security firm demonstrated they could bypass anti-malware AI using a similar approach to the earlier black box attack on the image classifier, but super-powered with an AI of their own.
Their system fed malicious code to the antivirus software and then noted the score it was given. It then used genetic algorithms to iteratively tweak the code until it was able to bypass the defenses while maintaining its function.
All the approaches noted so far are focused on tricking pre-trained machine learning systems, but another approach of major concern to the cybersecurity industry is that of “data poisoning.” This is the idea that introducing false data into a machine learning system’s training set will cause it to start misclassifying things.
This could be particularly challenging for things like anti-malware systems that are constantly being updated to take into account new viruses. A related approach bombards systems with data designed to generate false positives so the defenders recalibrate their systems in a way that then allows the attackers to sneak in.
How likely it is that these approaches will be used in the wild will depend on the potential reward and the sophistication of the attackers. Most of the techniques described above require high levels of domain expertise, but it’s becoming ever easier to access training materials and tools for machine learning.
Simpler versions of machine learning have been at the heart of email spam filters for years, and spammers have developed a host of innovative workarounds to circumvent them. As machine learning and AI increasingly embed themselves in our lives, the rewards for learning how to trick them will likely outweigh the costs.
Image Credit: Nejron Photo / Shutterstock.com Continue reading
Posted in Human Robots
#430761 How Robots Are Getting Better at Making ...
The multiverse of science fiction is populated by robots that are indistinguishable from humans. They are usually smarter, faster, and stronger than us. They seem capable of doing any job imaginable, from piloting a starship and battling alien invaders to taking out the trash and cooking a gourmet meal.
The reality, of course, is far from fantasy. Aside from industrial settings, robots have yet to meet The Jetsons. The robots the public are exposed to seem little more than over-sized plastic toys, pre-programmed to perform a set of tasks without the ability to interact meaningfully with their environment or their creators.
To paraphrase PayPal co-founder and tech entrepreneur Peter Thiel, we wanted cool robots, instead we got 140 characters and Flippy the burger bot. But scientists are making progress to empower robots with the ability to see and respond to their surroundings just like humans.
Some of the latest developments in that arena were presented this month at the annual Robotics: Science and Systems Conference in Cambridge, Massachusetts. The papers drilled down into topics that ranged from how to make robots more conversational and help them understand language ambiguities to helping them see and navigate through complex spaces.
Improved Vision
Ben Burchfiel, a graduate student at Duke University, and his thesis advisor George Konidaris, an assistant professor of computer science at Brown University, developed an algorithm to enable machines to see the world more like humans.
In the paper, Burchfiel and Konidaris demonstrate how they can teach robots to identify and possibly manipulate three-dimensional objects even when they might be obscured or sitting in unfamiliar positions, such as a teapot that has been tipped over.
The researchers trained their algorithm by feeding it 3D scans of about 4,000 common household items such as beds, chairs, tables, and even toilets. They then tested its ability to identify about 900 new 3D objects just from a bird’s eye view. The algorithm made the right guess 75 percent of the time versus a success rate of about 50 percent for other computer vision techniques.
In an email interview with Singularity Hub, Burchfiel notes his research is not the first to train machines on 3D object classification. How their approach differs is that they confine the space in which the robot learns to classify the objects.
“Imagine the space of all possible objects,” Burchfiel explains. “That is to say, imagine you had tiny Legos, and I told you [that] you could stick them together any way you wanted, just build me an object. You have a huge number of objects you could make!”
The infinite possibilities could result in an object no human or machine might recognize.
To address that problem, the researchers had their algorithm find a more restricted space that would host the objects it wants to classify. “By working in this restricted space—mathematically we call it a subspace—we greatly simplify our task of classification. It is the finding of this space that sets us apart from previous approaches.”
Following Directions
Meanwhile, a pair of undergraduate students at Brown University figured out a way to teach robots to understand directions better, even at varying degrees of abstraction.
The research, led by Dilip Arumugam and Siddharth Karamcheti, addressed how to train a robot to understand nuances of natural language and then follow instructions correctly and efficiently.
“The problem is that commands can have different levels of abstraction, and that can cause a robot to plan its actions inefficiently or fail to complete the task at all,” says Arumugam in a press release.
In this project, the young researchers crowdsourced instructions for moving a virtual robot through an online domain. The space consisted of several rooms and a chair, which the robot was told to manipulate from one place to another. The volunteers gave various commands to the robot, ranging from general (“take the chair to the blue room”) to step-by-step instructions.
The researchers then used the database of spoken instructions to teach their system to understand the kinds of words used in different levels of language. The machine learned to not only follow instructions but to recognize the level of abstraction. That was key to kickstart its problem-solving abilities to tackle the job in the most appropriate way.
The research eventually moved from virtual pixels to a real place, using a Roomba-like robot that was able to respond to instructions within one second 90 percent of the time. Conversely, when unable to identify the specificity of the task, it took the robot 20 or more seconds to plan a task about 50 percent of the time.
One application of this new machine-learning technique referenced in the paper is a robot worker in a warehouse setting, but there are many fields that could benefit from a more versatile machine capable of moving seamlessly between small-scale operations and generalized tasks.
“Other areas that could possibly benefit from such a system include things from autonomous vehicles… to assistive robotics, all the way to medical robotics,” says Karamcheti, responding to a question by email from Singularity Hub.
More to Come
These achievements are yet another step toward creating robots that see, listen, and act more like humans. But don’t expect Disney to build a real-life Westworld next to Toon Town anytime soon.
“I think we’re a long way off from human-level communication,” Karamcheti says. “There are so many problems preventing our learning models from getting to that point, from seemingly simple questions like how to deal with words never seen before, to harder, more complicated questions like how to resolve the ambiguities inherent in language, including idiomatic or metaphorical speech.”
Even relatively verbose chatbots can run out of things to say, Karamcheti notes, as the conversation becomes more complex.
The same goes for human vision, according to Burchfiel.
While deep learning techniques have dramatically improved pattern matching—Google can find just about any picture of a cat—there’s more to human eyesight than, well, meets the eye.
“There are two big areas where I think perception has a long way to go: inductive bias and formal reasoning,” Burchfiel says.
The former is essentially all of the contextual knowledge people use to help them reason, he explains. Burchfiel uses the example of a puddle in the street. People are conditioned or biased to assume it’s a puddle of water rather than a patch of glass, for instance.
“This sort of bias is why we see faces in clouds; we have strong inductive bias helping us identify faces,” he says. “While it sounds simple at first, it powers much of what we do. Humans have a very intuitive understanding of what they expect to see, [and] it makes perception much easier.”
Formal reasoning is equally important. A machine can use deep learning, in Burchfiel’s example, to figure out the direction any river flows once it understands that water runs downhill. But it’s not yet capable of applying the sort of human reasoning that would allow us to transfer that knowledge to an alien setting, such as figuring out how water moves through a plumbing system on Mars.
“Much work was done in decades past on this sort of formal reasoning… but we have yet to figure out how to merge it with standard machine-learning methods to create a seamless system that is useful in the actual physical world.”
Robots still have a lot to learn about being human, which should make us feel good that we’re still by far the most complex machines on the planet.
Image Credit: Alex Knight via Unsplash Continue reading
Posted in Human Robots