The network architecture and parameters used for both the injection (GANinj) and removal (GANrem) networks. Source: arXivThe network architecture and parameters used for both the injection (GANinj) and removal (GANrem) networks. Source: arXiv

Cybersecurity risks pose threats to a growing number of areas, such as interference in elections and stealing sensitive corporate data. Now add medical imagery as a target: Malware can be designed to tamper with medical scans for a variety of nefarious reasons, according to researchers from Ben-Gurion University and Soroka University Medical Center, Israel.

The team designed such software using deep learning artificial intelligence to determine the potential to tamper with CT and MRI scanning equipment and produce false lung cancer results. Security weaknesses were found both in medical imaging equipment and networks transmitting those images, and the malware was demonstrated to either add fake cancerous growths to images or to obscure real lesions and nodules. Motivations for such actions could include insurance fraud, political influence or other possible goals.

The malware altered 70 images sufficiently enough to convince three radiologists into believing patients had cancer. When viewing scans with fabricated cancerous nodules, the radiologists diagnosed cancer 99% of the time. In cases where real cancerous nodules were removed from scans, the radiologists maintained that those patients were healthy 94% of the time.

The researchers used GAN deep learning technology, which consists of two neural networks: The generator and discriminator. The system generates new images that are visually similar to real ones comprising a sample data distribution. Security countermeasures recommended in the study, which is published in arXiv, include the addition of digital watermarks to images and continually updating servers and anti-virus software on radiologist workstations.

To contact the author of this article, email shimmelstein@globalspec.com