How to Spot and Prevent Deepfakes Spreading Medical Misinformation 

Back in 2021, many people died and the US economy suffered losses worth $50-300 million each day – both from people refusing to get COVID-19 vaccines due to the spread of medical misinformation.

If this is the damage that social media conspiracy theories alone can cause, imagine what will happen once more sophisticated technology like deepfakes enters the mix. With deep learning algorithms being used to manufacture falsified healthcare videos at scale, the potential for damage can be far greater.

Christopher Ross, of the RAND Corporation, recently led a study on the damage potential of scientific and medical deepfakes. The group found that 27-50% of people across different populations are completely unable to distinguish between the real videos and the fake ones.

“Imagine a scenario where the likeness of well-respected doctors is deepfaked to endorse questionable health products online,” warns Justin Marciano, CEO of a deepfake detection startup called IdentifAI. “While these fabricated endorsements accumulate millions of views, the implications are disastrous for the medical community.” We are already witnessing fake endorsements from deepfakes of celebrities. It is only a matter of time that this spreads to healthcare.

Deepfakes could also be used to spread false claims against pharmaceutical products, manipulate medical research data, or impersonate healthcare company executives to commit identity theft.

As a healthcare, biotech, or pharma company — how do you insulate your organization against these pervasive threats? At Eularis, we have helped companies like Pfizer, Roche, and Sanofi use strategic employee training and business consultation to understand and operationalize the impact of Deepfakes. Here’s how we recommend preventing and combatting deepfake attacks against your business.

What are Deepfakes?

A deepfake is a type of synthetic media where a person’s face or voice is cloned by another using AI, making it appear as though they are saying or doing things that never actually happened. This is achieved through advanced deep learning networks and algorithms.

Imagine you have a jigsaw puzzle, where each piece represents a data point about a person’s facial expressions, voice, or movements. When creating a deepfake, AI algorithms, like deep learning networks, work to understand and replicate how these pieces fit together to represent someone’s identity.

• The first step is gathering visual or audio data of the individual involved
• This data is fed into sophisticated AI models, typically Generative Adversarial Networks (GANs). They create the deepfake content, then verify its authenticity against real data. This is repeated several times until the AI models are properly trained.
• The last step often involves refining the output to remove any glitches or unnatural artefacts that could give away the fact that it’s a fake.

While it is possible to harness deepfake technology for positive impact, the threat of it being used for business identity theft, corporate phishing, and more — is very real.

In fact, this is already being done. In a report from CNN, the Hong Kong police recently revealed how deepfakes were used to lure a business employee into remitting $25.6 million USD to con artists. The worker was tricked using a multi-person video conference that appeared to include several high-level company executives, including the CFO — all of whom were deepfakes. How long before such attacks affect the healthcare and pharmaceutical industry, too?

The challenge today is anyone can create a deep fake of anyone with only a few seconds of video images of that person that can be picked up from social media but the damage that can be caused is significant.

Deepfakes Vs Digital Twins

We see a lot of confusion between deepfakes with digital twins. But there’s a clear difference between to the two concepts:

A deepfake is an artificial image, video, or audio recording that has been convincingly edited, using deep learning ( type of machine learning AI) algorithms, to depict a real person saying or doing something they did not actually say or do. They are often created by malicious actors without the consent of the person being depicted, with a clear intent to cause harm.

A digital twin is a very different thing – it is a virtual model or representation of a real-world object, system, or process. It is designed to accurately simulate the characteristics and behavior of its physical counterpart. Human digital twins are extremely complex as they replicate the person’s body processes and they are highly accurate. So much so that the FDA changed their regulations in 2022 to allow digital twins to be used in lieu of humans in clinical trials. Digital twins are built using real-time data from sensors to keep the virtual and physical versions in sync. So a simple example is a digital twin of a person’s pancreas. The base model of the human pancreas created from thousands of human’s data is the initial model. This then gets adapted to the individual who is using that insulin pump. The sensor continuously sends data about insulin levels to the insulin pump containing the digital twin data and that runs a mathematical model of glucose metabolism. The model is calibrated to the patient’s blood glucose readings, health status and individual characteristics, such as gender, age, weight, and activity level to mimic that person’s pancreas.

How to Spot a Deepfake?

“You might think that as deepfakes proliferate, people are going to get good at it just by being able to pick it out better with experience,” says Christopher Doss. But he warns that the opposite might be true.

Spotting a deepfake requires a keen eye for detail. In fact, your brain may very well register that there’s something wrong long before you consciously realize it. Still, here are some things that could give it away – although they get better and better constantly:

• Unnatural Eye Movement and Lack of Blinking: Deepfakes often struggle to replicate the natural movement of eyes or the act of blinking convincingly. A consistent absence of blinking or unnatural eye movements could be a significant red flag.
• Awkward Facial Expressions and Misalignments: If facial expressions seem off or don’t match the emotional tone of the speech, or if there’s awkward positioning of facial features (like the nose not aligning with the direction the face is pointing), these could indicate a deepfake.
• Inconsistent Audio: Deepfakes often focus more on the visual aspects, leading to discrepancies in audio quality. Misalignments between the voice and lip movements, robotic-sounding voices, or unnatural pronunciations are common giveaways. Additionally, digital noise or the absence of natural background sounds might be present.
• Unnatural Coloring and Lighting: Look for abnormal skin tones, discoloration, strange lighting, or misplaced shadows. These visual anomalies can suggest that the footage has been digitally altered.
• Poorly Rendered Hair and Teeth: Deepfake algorithms may not accurately replicate the individual characteristics of hair and teeth. Lack of detailed textures, such as frizzy hair or defined teeth edges, can be clues.
• Blurring or Misalignment at Edges: Noticeable blurring where the face meets the neck, or misaligned visuals, particularly in the hairline or where different parts of the body come together, are potential indicators of a deepfake.
• Unnatural Body Movement or Posture: Deepfakes primarily focus on the face, often neglecting the natural movement or positioning of the body. If the body shape appears distorted, or if there’s awkward or inconsistent positioning, it might be a deepfake.
• Reflections and Environment Consistency: Inconsistencies in reflections in the eyes or mismatching environmental lighting can also reveal a deepfake. Natural reflections should mirror the surroundings and match across both eyes.