Trusted Enterprise AI and ML

A Point of View

While machine learning has been around for quite some time, the release of Chat GPT in November 2022 has generated significant interest among consumers and enterprises. The cognitive and text-generation capabilities of Chat GPT have sparked discussions about the social impact of AI, as well as concerns related to security, privacy, and copyright issues. Responsible and ethical adoption of this powerful technology is now a priority.

Enterprise CEOs, keen not to be left behind, are urging their departments to innovate and explore use cases for generative AI. Enterprises are leveraging generative AI and machine learning to gain insights, summarize vast amounts of information, generate reports (including sales forecasts, annual reports, and ESG initiatives), negotiate contracts, review legal documents, make predictions, and classify items. In many instances, they take actionable steps based on the recommendations and insights provided.

A Conceptual Enterprise AIML Layer

In the accompanying graphic, we present a simplified conceptual view of the enterprise stack. Enterprise data comprises information acquired from external partners, sourced from third-party data providers, and internally generated through various applications and content creation tools. These diverse data sources are carefully curated and serve as the foundation for training models.

The stack itself is intentionally simplified, and while there are multiple ways to depict it, we’ll focus on the essential components. At the base, we find pre-trained language models (LLMs) and vector databases — either purpose-built or derived from open-source or vendor-specific solutions. These foundational resources support higher-level enterprise-specific AI and ML models. Additionally, the stack includes agents, prompt libraries, and other reasoning objects.

A critical layer in this ecosystem is the trust layer, provided by blockchain or distributed ledger technology (DLT). Blockchain’s immutability, distributed ledger capabilities, and support for smart contracts make it a valuable choice. However, other DLTs, such as QLDB or Fluree, can also be equally effective.

Actors, Attributes and Purpose

In an AI/ML solution, various actors play crucial roles. For simplicity, we’ll focus on three key actors: Users, Models (including Agents), and Prompts (both dynamic and static). Static prompts are predefined during the design phase for specific purposes.

Users:

• Users can be individuals or applications.
• Their primary role is to invoke Models and Agents.

Users interact with the system, seeking insights or performing specific tasks.

Models and Agents:

These represent the foundational domain-specific AI applications.

• Models are trained algorithms that perform specific tasks (e.g., image recognition, language translation).
• Agents act as intermediaries, facilitating communication between Users and Models.

Together, they enable the AI system to process requests and generate responses.

Prompts:

• Prompts serve as user or application-provided reasoning and queries.
• They guide the system by framing questions or expressing requirements.
• Prompts can be either dynamic (generated on-the-fly) or static (predefined).

Attributes of Actors:

Identity:

• Represented by cryptographic credentials.
• Used for authentication and authorization.

Role(s):

• Defines the functions an actor performs (e.g., User, Model Trainer, Administrator).

Access to Resources:

• Users have access to Models, Agents, and Prompts.
• Proper access control ensures security and privacy.

Challenges:

Prompts Filtering:

• Prompts must be filtered to remove spurious or malicious content.
• Sensitive information (PII) should be handled carefully.

Response Filtering:

• Responses generated by Models need filtering.
• Avoid sensitive, incorrect, or harmful information.

Overall, managing these actors and their attributes ensures responsible and effective AI/ML system operation.

PII (Personally Identifiable Information):

• PII refers to any information that can be used to identify an individual. Examples include names, addresses, social security numbers, and email addresses.
• In the context of AI and data handling, safeguarding PII is crucial to protect privacy and prevent misuse.

Prompts:

• Prompts are user or application-provided instructions or queries.
• When using prompts, it’s essential to filter out any inappropriate content, including examples that exclude specific populations.

Models:

• Models are AI functions that perform specific tasks based on their training.
• They leverage knowledge from a pre-existing dataset (often stored in a vector database) and apply it to generate insights or outcomes.
• Language models (LLMs) play a significant role in understanding and generating text.

Remember, responsible AI development involves thoughtful handling of data, ethical considerations, and ensuring that the technology benefits everyone without causing harm.

DLT Enabled Enterprise AI Trust Protocol

In the realm of AI and ML, robust governance is essential to ensure responsible usage and maintain auditability, especially in cases of unexpected behavior or outcomes. Users must have confidence in the models they employ and trust the insights these models generate. Let’s delve into the components of a DLT-enabled trust protocol:

User Trust and Model Assurance:

• User Trust: Users should trust the models they interact with. This trust extends to the outcomes and insights provided by these models.
• Model Assurance: Ensuring that models are reliable, accurate, and well-behaved is critical. Audit trails and transparency play a key role here.

Interaction Pattern:

The diagram illustrates an interaction pattern between AI/ML components. However, for brevity, we won’t delve into user identity and access management, which is a well-established area.

Model Registration on DLT:

• When a model is ready for deployment, it should be registered on the distributed ledger (DLT).
• Registration details may include model specifications, verification records, and information about the training data used.
• Successful validation and verification lead to the issuance of cryptographic credentials by the blockchain.

Model Lifecycle Management:

• The blockchain plays a crucial role in managing the entire lifecycle of models.
• This includes deployment, updates, and eventual retirement.
• Similarly, pre-defined prompts and agents can also benefit from DLT-based lifecycle management.

Runtime Authentication and Recording:

• During runtime, models must be authenticated and authorized.
• Data recording encompasses various aspects, such as startup state, data consumption digests, and access details for applications or users.

In Conclusion

Users’ trust in AI models and the reliability of generated insights are paramount. Distributed ledger technologies (DLTs) offer governance mechanisms that support responsible AI usage and facilitate model lifecycle management. A sample implementation model:

Chainyard provides practical consulting, advice, and implementation services for AI and ML use cases. Trust Your Supplier is a Chainyard solutions platform that supports enterprises for supplier qualification and risk management. The author welcomes comments, and feedback including errors and omissions.