Trade Secret Protection Of Proprietary Ai Model Architecture In Research Institutions.
1) What Are Trade Secrets & Why They Matter for AI Models
A trade secret is confidential information that gives a business (or research institution) a competitive or economic advantage and is actively protected as secret. In the context of AI:
✅ Source code
✅ Model architecture
✅ Training methods
✅ Hyperparameter configurations
✅ Data preprocessing techniques
✅ Loss functions or proprietary layers
can all be trade secrets — provided they are not publicly disclosed and are reasonably protected.
Elements of Trade Secret Protection (Uniform Trade Secrets Act & Defend Trade Secrets Act):
Information not generally known or readily ascertainable
Commercial value because it is secret
Reasonable efforts to keep it secret
In research institutions, trade secrets exist alongside publications and open research. The institution needs carefully drafted policies differentiating what is proprietary vs. publishable research.
2) Unique Challenges for AI Model Architecture
AI model architecture has features that raise special concerns:
Reverse Engineering Risk: Adversaries may extract architecture by querying a model’s API.
Collaborative Research: Research partners may inadvertently disclose parts of the model.
Publications vs Protection: Publishing research may destroy trade secret status.
Employee Mobility: Researchers often move between academia and industry.
So institutions must:
📌 Use stringent access controls.
📌 Classify inventions appropriately.
📌 Use NDAs and employee agreements.
3) Case Laws Illustrating Trade Secret Protection in AI & Related Technologies
Below are six detailed cases demonstrating how courts handled trade secret claims involving software, algorithms, and advanced technologies:
Case 1 — Waymo LLC v. Uber Technologies, Inc. (N.D. Cal. 2017)
Facts
Waymo (Alphabet/Google’s self‑driving car division) alleged that a former employee, Anthony Levandowski, downloaded thousands of confidential files about Waymo’s LIDAR and autonomous driving software before leaving to start a company that was then acquired by Uber.
Trade Secrets at Issue
LIDAR design and engineering documents
Source code and architectures relating to self‑driving technology
Simulation tools and proprietary testing software
Court’s Analysis
The court found that Waymo’s materials were protected as trade secrets because:
✔ Waymo restricted access to these files
✔ Waymo used internal labeling and access logs
✔ The files were not publicly known
Levandowski’s systematic download was found to violate Waymo’s trade secret rights.
Outcome
Uber agreed to:
➡ Give Waymo 0.34% of Uber’s equity (~$245 million)
➡ Respect Waymo’s intellectual property rights
➡ Not use Waymo’s confidential materials
Relevance to AI Models
If an AI architecture is properly restricted and confidential, courts will enforce trade secret protection even when:
📍 The person is a former employee
📍 The subsequent technology is similar
Case 2 — Google LLC v. Oracle America, Inc. (Supreme Court of the United States, 2021)
Not a pure trade secret case, but deeply relevant to AI model protection
Oracle argued Google unlawfully used Java APIs in Android. Google argued compatibility and fair use.
While primarily a copyright fair use case, the opinions acknowledged how API structure, organization, and interface logic can have commercial value — a point relevant to AI architecture.
Key Takeaway
Software architecture and the organization of functional components can be proprietary even if not directly copyrighted, reinforcing the idea that:
📌 Architectural design can be protectable
📌 Similar reasoning supports confidentiality claims when architecture is secret
This case helps clarify how courts view software structure as valuable intellectual property.
Case 3 — Epic Systems Corporation v. Tata Consultancy Services (W.D. Wis. 2016)
Facts
Epic (electronic medical records software) sued TCS for stealing trade secrets through former employees and uploading source code to cloud storage later accessed by TCS.
Court’s Findings
Software code and related systems were trade secrets
Employees improperly uploaded confidential materials to cloud storage before leaving
TCS gained access and used that code
Outcome
Court held in favor of Epic, finding:
➡ Misappropriation by former employees
➡ TCS received and used trade secret materials
AI Relevance
This highlights that cloud access without proper authorization still constitutes misappropriation, even if the receiver claims no intent.
Case 4 — DuPont v. Christopher (E.D. Va. 2009)
Facts
A DuPont researcher downloaded confidential materials on supercement formulation and shared them with another company.
Why It Matters
The court identified:
✔ DuPont had proper secrecy safeguards
✔ The formulation was economically valuable
✔ There was clear evidence of unauthorized access and removal
Trade Secret Law Principles Reinforced
📍 Actual usage or dissemination is enforcement evidence
📍 Ownership and maintenance of secrecy are crucial
Case 5 — Kewanee Oil Co. v. Bicron Corp. (U.S. Supreme Court 1974)
Historic Foundation Case
Facts
Kewanee sued Bicron alleging misappropriation of trade secrets involving chemical processes.
Supreme Court Holding
State trade secret protection does not violate federal patent policy, even if the invention might be patentable.
Why It Matters
This firmly establishes that trade secret rights are independent and strong, especially for technologies where patenting is impractical—like fast‑evolving AI models.
Case 6 — IBM v. Visentin (N.D. Cal. 2010)
Facts
Former IBM IT employee took confidential IBM documents before starting at another company.
Court’s Decision
Ordered:
➡ Preliminary injunction to stop use
➡ Employee barred from working on projects that could use stolen trade secrets
Principle
Courts can impose remedies beyond damages when:
📌 There is real threat of irreparable harm
📌 Secrecy violations could significantly disadvantage the plaintiff
Case 7 — Thermo King v. Whiteford (Minn. Ct. App. 1984)
AI/Software Analogy
Although older, this case holds that taking internal manuals and confidential designs can constitute misappropriation even without evidence that defendant used the information.
Lesson
Trade secret laws focus on the act of taking and failing to protect information, not merely its use.
4) General Guidance for Protecting AI Architecture in Research Institutions
To ensure robust trade secret protection:
A. Classification & Policies
Define what counts as proprietary: model architectures, meta‑parameter tuning processes, training pipelines.
Separate research intended for publication from proprietary work.
B. Access Controls
Use role‑based access
Log usage
Monitor downloads
Courts consistently emphasize reasonable efforts to maintain secrecy as a factor that wins trade secret status.
5) Practical Techniques for AI Models
| Technique | Why It Helps Trade Secret Protection |
|---|---|
| API‑only access to models | Prevents reverse engineering of architecture |
| Encoded model outputs vs source transparency | Avoids disclosure of internal structures |
| Differential privacy/Secure enclaves | Technical barrier to extracting secrets |
| NDAs with collaborators | Creates enforceable legal commitments |
| Compartmentalization of teams | Limits who knows architecture details |
6) Trade Secret vs Patent — Which for AI?
📌 Patent
Protects specific inventions
Public disclosure required
Strong exclusionary rights
📌 Trade Secret
Protects confidential know‑how
No expiry so long as secret is kept
Does not prevent independent discovery
For AI architectures that are hard to reverse engineer but not easily described in patent claims, trade secret protection may be more practical.
7) Enforcement Considerations Specific to AI
Reverse engineering defenses may be weaker for models exposed publicly.
APIs that can be queried may leak architectural structure.
Courts may balance:
Economic value of the secret
Efforts to maintain secrecy
How widely technology is exposed
If a model is fully open (weights + code), trade secret protection is lost.
8) Key Takeaways for Research Institutions
Treat proprietary AI architectures as trade secrets if not meant for publication.
Implement documented safeguards — courts focus on whether “reasonable efforts” were made.
Use legal agreements (NDAs, IP assignment) to protect rights.
Monitor access and have incident plans for suspected theft.
Enforce rights promptly — delays can weaken claims.
RELATED Blog
comments