IP Concerns For Machine Learning Fraud Detection In Vietnamese Banks
1. Introduction
Machine Learning-based Fraud Detection Systems are increasingly used by banks in Vietnam to:
Detect suspicious transactions
Identify money laundering
Prevent phishing or card fraud
Monitor anomalous customer behavior
These systems rely on advanced algorithms, proprietary software, and large datasets of banking transactions.
While these tools improve security and efficiency, they raise several IP-related issues:
Ownership of ML models
Copyright protection of algorithms and code
Patentability of fraud detection methods
Trade secret protection vs. transparency
Dataset licensing and privacy considerations
Courts globally have addressed analogous issues in software, AI, and financial technology, which are directly relevant to Vietnamese banks.
2. Key IP Concerns
(A) Ownership of Machine Learning Models
ML models are often developed by:
Banks internally
External vendors or fintech companies
Questions arise:
Who owns the resulting models?
Can banks claim IP if developed by third-party vendors?
Are AI-generated detection patterns copyrightable or patentable?
(B) Copyright in Code and Software
Fraud detection models include:
Source code
Preprocessing pipelines
Feature extraction scripts
Copying or adapting code without license could lead to copyright infringement, even internally if licensing agreements are violated.
(C) Patent Issues
Some banks or vendors may attempt to patent ML methods for fraud detection.
Legal challenges arise if algorithms are considered abstract ideas or mere mathematical formulas.
(D) Trade Secrets vs. Transparency
ML models often rely on proprietary scoring algorithms.
Banks may claim trade secret protection for:
Model architecture
Feature importance rankings
Detection thresholds
However, regulators may demand explainability and auditability, creating tension with IP protection.
(E) Dataset Ownership
ML fraud detection requires historical transaction datasets.
Issues arise when datasets are:
Shared with third-party vendors
Derived from multiple banks
Proprietary or confidential
Unauthorized use may violate copyright, licensing, or banking secrecy laws.
3. Relevant Case Laws
Here are seven key cases illustrating IP issues applicable to ML fraud detection:
Case 1: Alice Corp. v. CLS Bank International (2014)
Court
United States Supreme Court
Facts
Alice Corp held patents for a computer-implemented method for reducing financial risk. CLS Bank argued that the patent claimed an abstract idea.
Issue
Are computer-implemented algorithms patentable?
Judgment
Abstract ideas implemented on a computer are not patentable
There must be an inventive concept beyond an abstract idea
Relevance to ML Fraud Detection
Simple ML algorithms for fraud detection may not be patentable unless they demonstrate novel technical implementation
Banks attempting patents for detection rules alone may fail
Case 2: Feist Publications v. Rural Telephone Service (1991)
Court
United States Supreme Court
Facts
Feist copied telephone listings from Rural Telephone to create its own directory.
Judgment
Facts are not copyrightable
Only creative selection or arrangement of data is protected
Relevance
Transaction data itself (e.g., account numbers, timestamps) is not copyrightable
ML models can use these raw facts without infringement
However, copying structured datasets or feature-engineered tables may raise copyright issues
Case 3: Google LLC v. Oracle America, Inc. (2021)
Court
United States Supreme Court
Facts
Google used portions of Oracle’s Java APIs to build Android.
Judgment
Using APIs for interoperability is fair use
Relevance
ML fraud detection systems often integrate with banking software APIs
Copying functional API calls for interoperability is likely legal
Copying the entire proprietary code is not
Case 4: SAS Institute Inc. v. World Programming Ltd (2013, CJEU)
Court
Court of Justice of the European Union
Facts
World Programming developed software to run programs compatible with SAS software.
Judgment
Software functionality is not copyrightable
Only the source code is protected
Relevance
ML systems replicating functionality of existing fraud detection tools may be legal
Copying code directly remains infringement
Case 5: Thomson Reuters v. Ross Intelligence (2023)
Court
US District Court, Delaware
Facts
Ross Intelligence trained an AI on proprietary legal databases to provide legal research services.
Judgment
Training AI using proprietary datasets without authorization may constitute infringement
Relevance
ML fraud detection models in Vietnamese banks often train on historical transaction data
Proper authorization or licensing of data is required to avoid infringement
Case 6: Naruto v. Slater (Monkey Selfie, 2018)
Court
United States Court of Appeals, Ninth Circuit
Judgment
Non-human creators cannot hold copyright
Relevance
If ML autonomously detects fraud or generates reports:
Outputs may not be copyrighted
Ownership may reside with the bank or human supervisors
Case 7: Sega Enterprises Ltd. v. Accolade Inc. (1992)
Court
United States Court of Appeals, Ninth Circuit
Facts
Accolade reverse engineered Sega consoles to make compatible games.
Judgment
Reverse engineering for interoperability is fair use
Relevance
ML fraud detection may require reverse engineering banking protocols or formats
Legal if done for analysis, integration, or compliance, not copying proprietary algorithms
4. Additional IP Risks
Trade Secret Misappropriation
Third-party vendors could replicate proprietary detection logic
Data Licensing Violations
Sharing customer datasets across banks without consent may infringe IP rights
Patent Risk
Attempting to patent standard fraud detection techniques may fail due to abstract idea doctrine
Regulatory Transparency vs IP Protection
Regulators may require explainable AI, conflicting with secrecy claims
5. Legal Mitigation Strategies
Obtain proper licenses for proprietary software and datasets
Document human involvement in training and deployment to establish IP ownership
Protect trade secrets of ML architecture while meeting regulatory explainability
Avoid attempting patents for standard ML techniques unless genuinely novel
Use anonymized or open datasets for model training when possible
6. Conclusion
Machine learning fraud detection in Vietnamese banks intersects with IP concerns:
Copyright – source code, datasets, AI-generated outputs
Patents – abstract ML methods are generally not patentable
Trade secrets – detection algorithms vs. regulatory transparency
Dataset licensing – unauthorized use can lead to legal liability
Key Takeaways from Case Law:
Raw transactional data is free to use (Feist)
Transformative integration of software is fair use (Oracle v. Google)
Software functionality can be replicated legally (SAS Institute)
AI-generated outputs without human authorship are not copyrighted (Naruto)
Unauthorized use of proprietary datasets can result in infringement (Ross Intelligence)
Reverse engineering for interoperability is often legal (Sega v. Accolade)
Abstract ML methods may not be patentable (Alice v. CLS)
Overall Recommendation: Banks must balance IP protection, regulatory compliance, and operational transparency while deploying ML-based fraud detection.
RELATED Blog
comments