01 · Title Hero
Designing Cybersecurity for AI in Regulated Environments
Lessons from FDA Section 524B
Why AI Security is a System Problem, Not a Model Problem
Chia-Sheng (Jason) Lin · National Yang Ming Chiao Tung University
- Narrative
- 20 anchored chapters
- Regulatory anchor
- FDA Section 524B
- Core shift
- From model security to full-stack trust
Jump to a chapter20 sections
- 01Title Hero
- 02Opening Contrast
- 03Real-World Incident
- 04Untouched Model Paradox
- 05FDA Section 524B
- 06Full Stack Paradigm
- 07Hardware and Kernel Fragility
- 08Protocol Frontier
- 09Local Privilege Escalation
- 10On-Premise Myth
- 11Federated Learning
- 12Leakage Paradox
- 13Trust, Not Just Accuracy
- 14LeakPro
- 15Future Question
- 16SaMD DNA and Threat Modeling
- 17Default-Deny Clinical Environment
- 18Patch Governance Tradeoff
- 19Patient Safety
- 20Closing Thesis
翻訳状況
本文は英語のみです
この長編の Web 講演ページは現在英語版のみです。共通ナビゲーションと周辺 UI は、選択した言語のままご覧いただけます。
02 · Opening Contrast
If an AI recommends the wrong movie, it is fine.
The threshold for harm changes completely when AI participates in clinical judgment.
Ordinary software miss
If an AI recommends the wrong movie, it is fine.
Clinical miss
But what if that AI is outlining a tumor?
Why medical AI is different
That is not a bug.
That is a medical incident.
03 · Real-World Incident
A Real-World Incident in TaiwanMacKay Memorial Hospital ransomware attack
Public hospital incidents remind us that the clinical system, not only the model artifact, determines safety.
Location
Taipei
Disruption
Medical IT infrastructure disrupted
System lesson
This was not an attack on the model.
It was an attack on the system.
Reported impact
500+ computers affected
04 · Untouched Model Paradox
The Untouched Model Paradox
Integrity checks on weights and checkpoints do not protect the workflow around them.
Model state
The model can remain untouched.
Clinical consequence
The system can still be compromised.
Model integrity does not guarantee clinical trust. The attack surface includes infrastructure, dependencies, orchestration, identity, and the human-facing output path.
05 · FDA Section 524B
FDA Section 524BIn healthcare, AI is not just software. It is clinical risk.
In regulated medical contexts, cybersecurity obligations sit alongside safety and effectiveness.
Requirement
SBOM
Requirement
Postmarket vulnerability monitoring
Requirement
Patching / update capability
Regulatory message
AI security is no longer a best practice. It is a legal requirement.
06 · Full Stack Paradigm
Defending the full stack: a new architectural paradigm
Risk moves across layers, so defense and accountability have to move with it.
Layer
Governance
Layer
Model
Layer
Runtime
Layer
Kernel
Layer
Hardware
AI security is not only an engineering problem. It is also a governance problem, because trust depends on who can change the system, how evidence is collected, and whether operational controls survive deployment pressure.
07 · Hardware and Kernel Fragility
The foundation of fragility: hardware and kernel exploits
The lower layers define whether everything above them can still be trusted.
Foundation
Hardware
Foundation
Kernel
Trust impact
Flip a single bit. Collapse trust.
The most fragile layer is not always the model. A trusted model running on an untrusted foundation is still an untrusted clinical system.
08 · Protocol Frontier
The Protocol Frontier: exploiting frameworks and agents
Modern AI applications inherit the risks of orchestration, integration, and tool use.
Layer
Model
Layer
Agent
Layer
Tools
Layer
Connectors
Layer
Clinical systems
As AI systems gain tools, memory, and orchestration, the attack surface expands beyond the model. Security now has to account for workflows, service boundaries, and the connectors that bridge AI to clinical systems.
09 · Local Privilege Escalation
Anatomy of a local privilege escalation in Medical AI
A weak internal foothold can still distort the output path seen by clinicians.
Low-privilege foothold
Shared memory / queue access
Inference path tampering
Clinician-facing output impact
Safety note
Conceptual attack anatomy only.
10 · On-Premise Myth
If data never leaves the hospital, is it really safe?
Local hosting narrows some exposure, but it does not remove trust and access problems.
Hospital perimeter
On-premise does not equal trusted.
Inside the hospital does not equal secure.
Identity
Lateral movement
Insider risk
Misconfiguration
11 · Federated Learning
What is Federated Learning?
Federated learning changes where data lives, not whether the overall system needs threat modeling.
Local site
Hospital A
Local training stays inside the institution.
Local site
Hospital B
Local training stays inside the institution.
Local site
Hospital C
Local training stays inside the institution.
Coordinator flow
Raw patient data stays in each hospital. Only model updates move between hospitals and the coordinator.
12 · Leakage Paradox
Data may stay, but information can still leak
Information can leak through updates, gradients, or behavior even when raw records never move.
Distribution truth
Data stays local.
Security truth
Risk can still exist.
13 · Trust, Not Just Accuracy
The real problem was not accuracy. It was trust.
Accuracy scores alone do not restore confidence if the surrounding chain cannot be explained or controlled.
Trust chain
Data
Trust chain
Runtime
Trust chain
Output
Trust chain
Clinician
14 · LeakPro
LeakPro: test the model before you share it
Leakage testing should happen before a model crosses organizational boundaries.
Before sharing
Model ready for release?
Validation gate
Ask whether it leaks, not only whether it is accurate.
Pre-release leakage assessment provides evidence for governance, model sharing, and internal review. It turns privacy risk into something testable instead of assumed.
15 · Future Question
The Future Question for Medical AI
Medical AI is governed across time, not just at launch.
Stage 1
Build
Stage 2
Validate
Stage 3
Deploy
Stage 4
Monitor
Stage 5
Update
Stage 6
Revalidate
How do we govern systems that keep changing?
16 · SaMD DNA and Threat Modeling
Embedding security into the SaMD DNA
Threat modeling connects abstract architecture to concrete design and review decisions.
Threat model frame
Architecture becomes control points.
STRIDE
Trust boundary
Assets
Data flow
Control points
17 · Default-Deny Clinical Environment
The 'Default-Deny' clinical environment
A clinical environment benefits from explicit trust boundaries and permitted paths only.
Zone
Clinician workstation
Zone
AI inference enclave
Zone
Hospital core systems
Zone
Vendor access
Policy
Allowed paths only
Inside the hospital is not the same as trusted. A clinical architecture should restrict movement by default, make exceptions explicit, and treat vendor or integration access as bounded paths rather than ambient trust.
18 · Patch Governance Tradeoff
Balancing clinical stability with cybersecurity responsiveness
Security response has to respect both exploit urgency and validation reality.
Exploitability
Clinical impact
Validation burden
Rollback readiness
Decision principle
The right patch decision is not the fastest one. It is the most controllable one.
19 · Patient Safety
Patient safety supersedes cybersecurity speed.
Clinical systems are allowed to move carefully when patient harm is on the line.
Patient safety supersedes cybersecurity speed.
In clinical environments, patching must be fast enough, but also safe enough to trust.
20 · Closing Thesis
AI security is not about protecting models. It is about protecting the entire computing stack.
The endpoint of this talk is a shift from model-centric security to system-centric trust.
AI security is not about protecting models. It is about protecting the entire computing stack.
From model security to full-stack trust