Securing the Future of Healthcare Data with AI: My Journey to Stop Cyber Threats Before They Harm Lives
By Dr. Yasmin Makki Mohialden, with contributions from Saba
Abdulbaqi Salman, Maad M. Mijwil, Nadia Mahmood Hussien, Mohammad Aljanabi,
Mostafa Abotaleb, Klodian Dhoska, and Pradeep Mishra
A World Where Healthcare and Technology Are One
Not long ago, medical records were kept in locked cabinets. Today,
they exist in cloud servers, hospital databases, and AI-driven applications.
Artificial Intelligence (AI) is transforming healthcare — from predicting
diseases to helping doctors make life-saving decisions faster.
But with great power comes great risk.
If the data that trains these AI systems is corrupted, the consequences can be
disastrous:
- Wrong
diagnoses given to patients.
- Delayed
treatments that cost lives.
- Loss of
trust in the healthcare system itself.
One of the most dangerous threats we face is a data poisoning
attack — a cyberattack where fake or harmful data is secretly inserted into the
information that AI learns from. This “poison” can cause AI systems to make
deadly mistakes.
Why I Took on This Challenge
When I began this research with my colleagues, we asked ourselves:
“How can we ensure that AI in healthcare remains safe, trustworthy, and
resilient — even when someone is trying to attack it?”
We knew the answer would require more than just one solution. It
would need a new kind of defense — one that protects data privacy, stops cyber
threats early, and works across different countries and hospitals without ever
moving sensitive patient data.
That vision became the foundation for our research: Enhancing
Security and Privacy in Healthcare with Generative AI-Based Detection and
Mitigation of Data Poisoning Attacks.
A Global Effort
This was not a journey I took alone.
Our team brought together expertise from Iraq, Russia, Albania, and India —
blending skills in artificial intelligence, cybersecurity, healthcare systems,
and data science.
We came from different cultures and disciplines, but we shared one
mission: protect the integrity of healthcare AI so it can save lives, not
endanger them.
The Core of Our Solution
We combined three powerful technologies to create a defense
framework:
1. Federated Learning — Training Without Sharing Data
In traditional AI training, all the data is collected in one
central location. In healthcare, this is risky because patient data can be
exposed or stolen during transfer.
With federated learning, the AI model learns directly from data
stored in each hospital or clinic — without that data ever leaving the
building. Only the “knowledge” gained (in the form of model updates) is sent to
a central system, keeping private information safe.
 |
| Figure 1 - Concept Diagram of Federated Learning Workflow |
2. Homomorphic Encryption — Calculations Without Unlocking Data
Even when AI needs to process data, we keep it encrypted the entire
time using homomorphic encryption. This means the AI can make calculations
without ever “seeing” the actual data.
It’s like having a chef who can cook a meal without opening the
sealed ingredients — the food is prepared without exposing what’s inside.
3. Autoencoder-Based Anomaly Detection — The AI That Spots Trouble
We also needed a way to spot poisoned or unusual data quickly. For
this, we used an autoencoder, a type of AI that learns what “normal” looks
like.
Once trained, if it encounters strange or corrupted data, the
difference becomes obvious. We measure this difference using Mean Squared Error
(MSE). If the error is too high, we flag it as suspicious.
How We Tested Our Framework
We began with synthetic healthcare datasets — realistic but
fictional patient data including:
- Age
- Gender
- BMI (Body
Mass Index)
- Blood
Pressure
- Cholesterol
Levels
- Health
status label
We then deliberately “poisoned” parts of the data by adding
Gaussian noise — mimicking the type of subtle attacks hackers might use.
Using federated learning, the model trained across multiple
simulated “hospitals” without centralizing the data. The data stayed encrypted,
and the autoencoder kept watch for anomalies.
The Results
The results were clear:
- The
framework successfully detected poisoned data in multiple scenarios.
- Privacy
was never compromised — no raw patient data was shared.
- The system
adapted well to different dataset sizes and variations.
What We Learned
Our method is promising, but there are challenges ahead:
- Performance
Speed — Homomorphic encryption is secure but computationally heavy, which
can be a hurdle for real-time hospital environments.
- Threshold
Precision — Setting the right anomaly detection sensitivity is key. Too
strict, and harmless data gets flagged; too loose, and attacks can slip
through.
Why This Research Matters
Healthcare is a prime target for cyberattacks:
- Medical
data is valuable on the black market.
- Disruption
in hospitals can cost lives.
- AI is
becoming mission-critical in diagnosis, treatment, and logistics.
By combining privacy-first training with proactive anomaly
detection, we created a defense that can be adapted beyond healthcare — into
finance, transportation, and national infrastructure.
The Human Side
This research is not just about AI and encryption — it’s about trust.
When a patient shares their medical history, they trust that it
will be used only to help them. Doctors trust that the AI tools they use will
be accurate.
Our mission was to ensure that this trust is never broken by
invisible cyber threats.
Acknowledging the Team
While I present this under my Nordic R&D Bridge profile, the
work belongs to our entire team:
- Saba
Abdulbaqi Salman
- Maad M.
Mijwil
- Nadia
Mahmood Hussien
- Mohammad
Aljanabi
- Mostafa
Abotaleb
- Klodian
Dhoska
- Pradeep
Mishra
Their combined expertise made this project possible.
The Road Ahead
Next steps for us include:
- Deploying
this system in real hospital networks.
- Optimizing
encryption speed for real-time operations.
- Extending
this framework to other critical sectors.
Cyber threats evolve every day — our defenses must evolve even
faster.
Final Thoughts
When people talk about AI in healthcare, they imagine futuristic
robots diagnosing diseases. But the real breakthrough is in protecting the data
and systems behind those robots.
The future of healthcare AI depends not just on what it can do —
but on whether we can trust it completely.
Our work is one step toward making that trust a reality.