Machine learning systems are pretty good at finding hidden correlations in data and using them to infer potentially compromising information about the people who generate that data: for example, researchers fed an ML system a bunch of Google Play reviews by reviewers whose locations were explicitly given in their Google Plus reviews; based on this, the model was able to predict the locations of other Google Play reviewers with about 44% accuracy.
For several years, I've been covering the bizarre phenomenon of "adversarial examples (AKA "adversarial preturbations"), these being often tiny changes to data than can cause machine-learning classifiers to totally misfire: imperceptible squeaks that make speech-to-text systems hallucinate phantom voices; or tiny shifts to a 3D image of a helicopter that makes image-classifiers hallucinate a rifle
Adversarial examples have torn into the robustness of machine-vision systems: it turns out that changing even a single well-placed pixel can confound otherwise reliable classifiers, and with the right tricks they can be made to reliably misclassify one thing as another or fail to notice an object altogether. — Read the rest
A team of researchers from Microsoft and Harvard's Berkman Center have published a taxonomy of "Failure Modes in Machine Learning," broken down into "Intentionally-Motivated Failures" and "Unintended Failures."
In TrojDRL: Trojan Attacks on Deep Reinforcement Learning Agents, a group of Boston University researchers demonstrate an attack on machine learning systems trained with "reinforcement learning" in which ML systems derive solutions to complex problems by iteratively trying multiple solutions.
Mac Pierce created a simple wearable to challenge facial recognition: do a little munging to an image of a face, print it on heat transfer paper, iron it onto see-through mosquito netting, slice, and affix to a billed cap — deploy it in the presence of facial recognition cameras and you'll be someone else. — Read the rest
Jonathan Zittrain (previously) is consistently a source of interesting insights that often arrive years ahead of their wider acceptance in tech, law, ethics and culture (2008's The Future of the Internet (and how to stop it) is surprisingly relevant 11 years later); in a new long essay on Medium (shorter version in the New Yorker), Zittrain examines the perils of the "intellectual debt" that we incur when we allow machine learning systems that make predictions whose rationale we don't understand, because without an underlying theory of those predictions, we can't know their limitations.
I was honored to be invited to contribute to the New York Times's excellent "Op-Eds From the Future" series (previously), with an op-ed called "I Shouldn't Have to Publish This in The New York Times," set in the near-future, in which we have decided to solve the problems of Big Tech by making them liable for what their users say and do, thus ushering in an era in which all our speech is vetted by algorithms that delete anything that looks like misinformation, harassment, copyright infringement, incitement to terrorism, etc — with the result that the only place where you can discuss anything of import is newspapers themselves.
An adversarial preturbation is a small, human-imperceptible change to a piece of data that flummoxes an otherwise well-behaved machine learning classifier: for example, there's a really accurate ML model that guesses which full-sized image corresponds to a small thumbnail, but if you change just one pixel in the thumbnail, the classifier stops working almost entirely.
Researchers from KU Leuven have published a paper showing how they can create a 40cm x 40cm "patch" that fools a convoluted neural network classifier that is otherwise a good tool for identifying humans into thinking that a person is not a person — something that could be used to defeat AI-based security camera systems. — Read the rest
Researchers from Tencent Keen Security Lab have published a report detailing their successful attacks on Tesla firmware, including remote control over the steering, and an adversarial example attack on the autopilot that confuses the car into driving into the oncoming traffic lane.
Last year, Princeton researchers revealed a powerful new ad-blocking technique: perceptual ad-blocking uses a machine-learning model trained on images of pages with the ads identified to make predictions about which page elements are ads to block and which parts are not.
Machine learning image classifiers use context clues to help understand the contents of a room, for example, if they manage to identify a dining-room table with a high degree of confidence, that can help resolve ambiguity about other objects nearby, identifying them as chairs.
Robot law pioneer Ryan Calo (previously) teamed up with U Washington computer science and law-school colleagues to write Is Tricking a Robot Hacking? — a University of Washington School of Law Research Paper.
A group of Chinese computer scientists from academia and industry have published a paper documenting a tool for fooling facial recognition software by shining hat-brim-mounted infrared LEDs on the user's face, projecting CCTV-visible, human-eye-invisible shapes designed to fool the face recognition software.
Many people worry that 3D printers will usher in an epidemic of untraceable "ghost guns," particularly guns that might evade some notional future gun control regime that emerges out of the current movement to put sensible, minimal curbs on guns, particularly anti-personnel guns.
Machine learning models use statistical analysis of historical data to predict future events: whether you are a good candidate for a loan, whether you will violate parole, or whether the thing in the road ahead is a stop sign or a moose.
In Partial Information Attacks on Real-world AI, a group of MIT computer science researchers report on their continuing work fooling Google's image-classifier, this time without any knowledge of how the classifier works.
Machine-learning-based image classifiers are vulnerable to "adversarial preturbations" where small, seemingly innocuous modifications to images (including very trivial ones) can totally confound them.
The Open AI researchers were intrigued by a claim that self-driving cars would be intrinsically hard to fool (tricking them into sudden braking maneuvers, say), because "they capture images from multiple scales, angles, perspectives, and the like."