 |
| AI that detects cancer can be fooled |
A new study shows that artificial intelligence programs that scan medical images for signs of cancer can be deceived by hackers and cyber-attacks.
Researchers have shown that computer software in X-ray imaging can add or remove evidence of cancer; These changes fooled AI tools and human radiologists.
This can lead to an incorrect diagnosis. AI programs that help screen mammograms can say the screening is healthy if there are signs of cancer, or falsely say it has cancer if the patient does not have cancer.
It is not yet known if this type of hack will occur in the real world. But this new study continues to add more research and shows that health organizations need to prepare for it.
Hospitals and medical facilities are increasingly becoming a target for hackers who launch cyberattacks. In most cases, these attacks steal patient data or shut down the company's IT system until the company pays the ransom.
Both types of attacks can harm patients by disrupting hospital operations and making it difficult for medical staff to provide quality care.
But experts are increasingly concerned about the potential for a direct attack on people's health.
For example, security researchers have shown that hackers can remotely break into insulin pumps connected to the Internet and dispense dangerous doses of drugs.
Hackers who can alter medical images and influence a diagnosis also fall into this category.
A research team from the University of Pittsburgh has developed a computer program that can produce x-rays that show no signs of cancer. It can make X-rays look cancerous and show no signs of cancer.
They placed the processed images through an artificial intelligence program that was trained to recognize the signs of breast cancer. They hired five radiologists to assess whether the images were real or fake.
Artificial intelligence can be deceived by hacked photos
About 70% of the manipulated images can be deceived by the program. That is, AI mistakenly assumes that processed images that appear to be cancer-free are cancer-free. He also said the photos, which were edited to look like she had cancer, showed signs of cancer.
Radiologists get to know the manipulated images better than others. The accuracy of selecting fake photos ranges between 29% and 71%.
Other studies have also shown that cyberattacks on medical images can lead to misdiagnosis. A group of cybersecurity researchers showed in 2019 that hackers can add or remove evidence of lung cancer from CT scans. These changes also misled human radiologists and artificial intelligence software.
There are no general or significant issues for such violations. However, there are several reasons why hackers might want to tamper with content such as X-rays or lung cancer scans.
Hackers may be interested in targeting some patients such as politicians. Or he may want to change the check to get a payment from the insurance company.
Hackers can also manipulate random images and refuse to stop the manipulation until the hospital pays the ransom.
Studies like this show that health organizations and those who build artificial intelligence models should be aware that hackers who tamper with medical scans are possible.
The study authors said the fitted model should be shown during training to teach them to recognize fakes. Radiologists may also need training to identify false images.