The artificial intelligence that detects cancer can be tricked
A new study has found that the programs of artificial intelligence who check medical images for evidence of cancer can be fooled by hacking and cyberattacks.
The researchers showed that the computer program could add or remove evidence of cancer from X-ray imaging. These changes fooled both the AI tool and human radiologists.
This leads to an erroneous diagnosis. An AI program that helps screen x-rays can say the exam is healthy when there are signs of cancer, or incorrectly say a patient has cancer when they don’t have cancer. .
Such penetrations are not yet known in the real world. But the new study adds to a growing body of research that suggests healthcare organizations need to be prepared.
Hospitals and healthcare facilities are increasingly targeted by hackers with cyberattacks. In most cases, these attacks steal patient data or shut down organization computer systems until the organization pays a ransom.
Both types of attacks can harm patients by impairing hospital operations and preventing healthcare workers from providing quality care.
But experts are also increasingly concerned about the potential for more direct attacks on people’s health.
Security researchers have shown that hackers can remotely break into Internet-connected insulin pumps and deliver dangerous doses of drugs, for example.
Hacks that can alter medical images and affect diagnosis also fall into this category.
A research team from the University of Pittsburgh designed a computer program that would perform x-rays without signs of Cancer looking cancerous. He may also do x-rays that appear to be cancerous and show no signs of cancer.
They submitted the manipulated images to a program ofartificial intelligence trained to detect signs of breast cancer. They asked five radiologists to decide if the images were real or fake.
Artificial intelligence can be tricked by hacked photos
About 70% of the manipulated images were able to trick this program. In other words, the AI incorrectly declared that images that had been manipulated to appear cancer-free were cancer-free. He also said photos manipulated to look like she had cancer showed evidence of cancer.
As for radiologists, some were better than others at detecting manipulated images. Accuracy in selecting fake photos ranged between 29% and 71%.
Other studies have also shown that a cyberattack on medical images can lead to misdiagnoses. And a team of cybersecurity researchers showed in 2019 that hackers could add or remove evidence of lung cancer from a scanner. These changes also fooled human radiologists and artificial intelligence software.
There have been no general or notable instances of such a violation. But there are several reasons a hacker might want to tamper with things like X-rays or lung cancer scans.
The hacker may be interested in targeting a specific patient, such as a political figure. Or he may want to change his checks to be paid by the insurance company.
Hackers can also randomly manipulate images and refuse to stop tampering with them until the hospital pays a ransom.
Studies like this show that healthcare organizations and people designing AI models need to be aware that a hack that changes medical scans is a possibility.
The study author said manipulated models should be shown during their training to teach them how to spot fake images. Radiologists may also need to be trained to identify false images.