By Cade Metz and Craig S. Smith - March 21, 2019
Last year, the Food and Drug Administration approved a device that can capture an image of your retina and automatically detect signs of diabetic blindness.
This new breed of artificial intelligence technology is rapidly spreading across the medical field, as scientists develop systems that can identify signs of illness and disease in a wide variety of images, from X-rays of the lungs to C.A.T. scans of the brain. These systems promise to help doctors evaluate patients more efficiently, and less expensively, than in the past.
Similar forms of artificial intelligence are likely to move beyond hospitals into the computer systems used by health care regulators, billing companies and insurance providers. Just as A.I. will help doctors check your eyes, lungs and other organs, it will help insurance providers determine reimbursement payments and policy fees.
Ideally, such systems would improve the efficiency of the health care system. But they may carry unintended consequences, a group of researchers at Harvard and M.I.T. warns.
In a paper published on Thursday in the journal Science, the researchers raise the prospect of “adversarial attacks” — manipulations that can change the behavior of A.I. systems using tiny pieces of digital data. By changing a few pixels on a lung scan, for instance, someone could fool an A.I. system into seeing an illness that is not really there, or not seeing one that is.
Software developers and regulators must consider such scenarios, as they build and evaluate A.I. technologies in the years to come, the authors argue. The concern is less that hackers might cause patients to be misdiagnosed, although that potential exists. More likely is that doctors, hospitals and other organizations could manipulate the A.I. in billing or insurance software in an effort to maximize the money coming their way.
Samuel Finlayson, a researcher at Harvard Medical School and M.I.T. and one of the authors of the paper, warned that because so much money changes hands across the health care industry, stakeholders are already bilking the system by subtly changing billing codes and other data in computer systems that track health care visits. A.I. could exacerbate the problem.
“The inherent ambiguity in medical information, coupled with often-competing financial incentives, allows for high-stakes decisions to swing on very subtle bits of information,” he said.
The new paper adds to a growing sense of concern about the possibility of such attacks, which could be aimed at everything from face recognition services and driverless cars to iris scanners and fingerprint readers.
An adversarial attack exploits a fundamental aspect of the way many A.I. systems are designed and built. Increasingly, A.I. is driven by neural networks, complex mathematical systems that learn tasks largely on their own by analyzing vast amounts of data.
By analyzing thousands of eye scans, for instance, a neural network can learn to detect signs of diabetic blindness. This “machine learning” happens on such an enormous scale — human behavior is defined by countless disparate pieces of data — that it can produce unexpected behavior of its own.
In 2016, a team at Carnegie Mellon used patterns printed on eyeglass frames to fool face-recognition systems into thinking the wearers were celebrities. When the researchers wore the frames, the systems mistook them for famous people, including Milla Jovovich and John Malkovich.
A group of Chinese researchers pulled a similar trick by projecting infrared light from the underside of a hat brim onto the face of whoever wore the hat. The light was invisible to the wearer, but it could trick a face-recognition system into thinking the wearer was, say, the musician Moby, who is Caucasian, rather than an Asian scientist.
Researchers have also warned that adversarial attacks could fool self-driving cars into seeing things that are not there. By making small changes to street signs, they have duped cars into detecting a yield sign instead of a stop sign.
Late last year, a team at N.Y.U.’s Tandon School of Engineering created virtual fingerprints capable of fooling fingerprint readers 22 percent of the time. In other words, 22 percent of all phones or PCs that used such readers potentially could be unlocked.
The implications are profound, given the increasing prevalence of biometric security devices and other A.I. systems. India has implemented the world’s largest fingerprint-based identity system, to distribute government stipends and services. Banks are introducing face-recognition access to A.T.M.s. Companies such as Waymo, which is owned by the same parent company as Google, are testing self-driving cars on public roads.
Now, Mr. Finlayson and his colleagues have raised the same alarm in the medical field: As regulators, insurance providers and billing companies begin using A.I. in their software systems, businesses can learn to game the underlying algorithms.
If an insurance company uses A.I. to evaluate medical scans, for instance, a hospital could manipulate scans in an effort to boost payouts. If regulators build A.I. systems to evaluate new technology, device makers could alter images and other data in an effort to trick the system into granting regulatory approval.
In their paper, the researchers demonstrated that, by changing a small number of pixels in an image of a benign skin lesion, a diagnostic A.I system could be tricked into identifying the lesion as malignant. Simply rotating the image could also have the same effect, they found.
Small changes to written descriptions of a patient’s condition also could alter an A.I. diagnosis: “Alcohol abuse” could produce a different diagnosis than “alcohol dependence,” and “lumbago” could produce a different diagnosis than “back pain."
In turn, changing such diagnoses one way or another could readily benefit the insurers and health care agencies that ultimately profit from them. Once A.I. is deeply rooted in the health care system, the researchers argue, business will gradually adopt behavior that brings in the most money.
The end result could harm patients, Mr. Finlayson said. Changes that doctors make to medical scans or other patient data in an effort to satisfy the A.I. used by insurance companies could end up on a patient’s permanent record and affect decisions down the road.
Already doctors, hospitals and other organizations sometimes manipulate the software systems that control the billions of dollars moving across the industry. Doctors, for instance, have subtly changed billing codes — for instance, describing a simple X-ray as a more complicated scan — in an effort to boost payouts.