| Internet-Draft | ibn-generative-ai | July 2025 |
| Cassara', et al. | Expires 8 January 2026 | [Page] |
This document describes how to specialize AI models in order to offer a scalable, efficient path to creating highly targeted generative models for Intent-Based Networking.¶
This note is to be removed before publishing as an RFC.¶
The latest revision of this draft can be found at https://giuseppefioccola.github.io/draft-cgfabk-nmrg-ibn-generative-ai/draft-cgfabk-nmrg-ibn-generative-ai.html. Status information for this document may be found at https://datatracker.ietf.org/doc/draft-cgfabk-nmrg-ibn-generative-ai/.¶
Discussion of this document takes place on the Network Management Research Group mailing list (mailto:nmrg@irtf.org), which is archived at https://mailarchive.ietf.org/arch/browse/nmrg. Subscribe at https://www.ietf.org/mailman/listinfo/nmrg/.¶
Source for this draft and an issue tracker can be found at https://github.com/giuseppefioccola/draft-cgfabk-nmrg-ibn-generative-ai.¶
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.¶
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."¶
This Internet-Draft will expire on 8 January 2026.¶
Copyright (c) 2025 IETF Trust and the persons identified as the document authors. All rights reserved.¶
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License.¶
This document describes how transfer learning techniques can be adopted to design generative AI specialized models for Intent-Based Networking (IBN). It also describes tools, such as Low-rank Adaptation (LoRA), for achieving efficient and scalable transfer learning in data networks for designing targeted generative models for Intent-Based Networking (IBN).¶
The objective of this document is define a framework for the rapid adaptation and composition of specialized knowledge, addressing the challenges identified in [AI-Challenges] and enabling the dynamic, multi-domain use cases of [IBN-UseCases].¶
Future work should focus on interoperability, governance, and autonomous management to fully realize the potential of this approach.¶
IBN represents a paradigm shift in network management, aiming to bridge the gap between business objectives and network configurations. Unlike traditional networking, which requires manual, device-level configurations, IBN allows operators to specify high-level intents (e.g., "ensure low latency for video traffic"), which the system then automatically translates, enforces, and continuously validates. Key references include [IBN-UseCases], which outlines IBN use cases across enterprise, data center, and service provider environments.¶
IBN fundamentally changes the role of network administrators by focusing on desired outcomes instead of manual command-line configurations. The system must be capable of not only translating intents into actionable policies but also of continuously validating whether those intents are met. This requires advanced mechanisms for policy enforcement, telemetry collection, and feedback loops that ensure alignment between high-level business goals and actual network performance. IBN can reduce operational complexity, improve network agility, and significantly lower the time required to deploy new services.¶
Generative AI, particularly Large Language Models (LLMs), can enhance IBN by automating intent parsing, policy generation, and network troubleshooting. LLMs can understand natural language intents, generate low-level configurations, and adapt policies in real-time. This aligns with the challenges presented in [AI-Challenges], which stresses the need for models that can dynamically adapt to context and scale efficiently.¶
Generative AI can extend its role to include auto-remediation, dynamic configuration adjustments, and the creation of context-aware policies based on real-time network conditions. These capabilities are critical in environments with rapidly shifting traffic patterns, such as 5G networks, multi-cloud environments, and large-scale SDN deployments. By leveraging fine-tuned generative models, IBN systems can become self-configuring, self-optimizing, and self-healing, moving closer to the vision of fully autonomous networks.¶
Generic LLMs often lack the precision required for domain-specific applications like IBN. Specialization is critical to improve accuracy, reduce inference time, and minimize resource consumption. LoRA provides an efficient method to fine-tune large models using minimal computational resources, enabling targeted model specialization for distinct network domains or intent categories.¶
Specialization ensures that the AI system correctly interprets and applies complex domain-specific policies, such as those for Quality of Service (QoS), security compliance, and multi-vendor interoperability. Without specialization, generic models risk generating suboptimal or even invalid configurations. LoRA-based specialization enables the rapid creation and deployment of these focused models, supporting agile, intent-driven network operations while maintaining the scalability and cost-effectiveness required for wide deployment across heterogeneous infrastructures.¶
Transfer learning enables pretrained models to adapt to specific tasks with significantly less data and computational resources. In the context of IBN, this approach offers a dual advantage: enhancing the efficiency of model training and improving the reliability of intent recognition and execution. Indeed, it avoids the need for building models from scratch, which is not only resource-intensive but also prone to overfitting due to limited labeled network-specific data. By fine-tuning models on domain-specific datasets, developers can quickly create robust models that interpret network intents with high accuracy. Furthermore, it facilitates cross-domain knowledge integration, allowing the model to generalize better across various networking environments and topologies. Additionally, continual learning mechanisms can be integrated into the LLM-based IBN framework to refine performance over time, incorporating feedback loops that learn from real-world interactions and user corrections.¶
In conclusion, transfer learning significantly boosts the viability of LLM-based solutions in IBN by offering a cost-effective, accurate, and scalable method for intent interpretation and execution. This fusion of LLMs with transfer learning not only propels the automation of network operations but also ensures that the AI systems remain adaptive, interpretable, and aligned with organizational objectives in dynamic digital infrastructures.¶
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here.¶
LLMs like GPT and BERT can process complex, unstructured inputs and generate contextually relevant outputs. In IBN, LLMs are used to translate high-level intents into vendor-agnostic policies and device-specific configurations. The integration of these models into network management systems is emerging as a critical research area, as outlined in [AI-Challenges].¶
An example flow in Figure 1 illustrates how an intent like "minimize latency for video traffic" is parsed by a generative model and translated into specific QoS policies.¶
+---------------------+
| User High-Level |
| Intent |
+---------------------+
|
v
+---------------------+
| Generative Model |
| (Intent Parser + |
| Policy Generator) |
+---------------------+
|
v
+-----------------------+
| Network Configuration |
+-----------------------+
LoRA [LoRA] introduces trainable low-rank matrices into the layers of pre-trained models, allowing fine-tuning with significantly fewer parameters. This approach reduces the storage and computational footprint, enabling deployment of specialized models even on resource-constrained systems.¶
IBN requires high domain accuracy. For example, a generic model may misinterpret a telecommunications-specific intent. Specializing models using LoRA ensures they are finely tuned to the nuances of networking language, policies, and protocols.¶
LoRA fine-tunes only small, low-rank matrices, leaving the rest of the model unchanged. For instance, adapting a generic LLM to focus on BGP policy generation could be achieved by training a small adapter on BGP-related datasets, reducing the compute requirements compared to full fine-tuning.¶
Unlike traditional fine-tuning, which adjusts the entire model, or pruning, which removes less important weights, LoRA introduces new parameters without overwriting existing knowledge. This allows multiple specialized LoRA adapters to coexist and be swapped or combined as needed.¶
Empirical studies show that LoRA adapters can reduce memory consumption by over 70% compared to full fine-tuning. For example, fine-tuning a model for IPv6 routing policy generation using LoRA reduces the required GPU VRAM from 16 GB to approximately 5 GB while maintaining performance.¶
A Hub is a structured repository where specialized adapters are stored, indexed, and shared. It enables efficient reuse and modularity. The architecture typically includes:¶
Adapters can be categorized by networking domain (e.g., security, QoS, routing) and intent granularity (e.g., traffic type, SLA constraint). Cross-indexing improves discoverability and composability.¶
Flow refers to the systematic combination of multiple adapters to form a new, composite model capable of handling complex, multi-domain intents.¶
Example: Combining an adapter for SLA enforcement with another for security compliance to address intents like "guarantee low latency for critical applications within zero-trust boundaries."¶
+-------------------------+
| Base LLM |
+-------------------------+
| | ... \
| | \
+---------+ +----------+ +----------+
| Adapter | | Adapter | | Adapter |
| QoS | | Security |...| j |
+---------+ +----------+ +----------+
\ | /
\ | ... /
+-------------------+
| Composite |
| Specialized |
| Model |
+-------------------+
A service provider managing both MPLS and SDN domains can fuse LoRA adapters for each to generate cross-domain policies automatically in response to high-level intents.¶
In-network telemetry provides real-time feedback on policy effectiveness, enabling further fine-tuning or adapter updates, aligning with the adaptive model requirements discussed in [AI-Challenges].¶
Drawing inspiration from the Analytics Logical Function (AnLF) and Model Training Logical Function (MTLF) architectures defined in 3GPP NWDAF, this section proposes a logical architecture to manage, store, and orchestrate AI models for Intent-Based Networking. The architecture introduces the following logical components:¶
Model Repository Function (MRF): Stores adapters and base models.¶
Model Training and Specialization Function (MTSF): Handles fine-tuning, evaluation, and versioning.¶
Model Fusion and Composition Function (MFCF): Manages flow processes to create composite models.¶
Intent Processing and Adapter Orchestration Function (IPOF): Dynamically selects, composes, and deploys adapters based on parsed intents.¶
Telemetry Feedback Function (TFF): Continuously monitors performance and triggers model re-specialization when required.¶
+----------------------------+
| Intent Interface |
+----------------------------+
|
v
+-------------------------------+
| Intent Processing and |
| Adapter Orchestration (IPOF) |
+-------------------------------+
|
v
+------------------------------+
| Model Fusion and Composition |
| (MFCF) |
+------------------------------+
|
v
+-----------------------------+
| Model Repository (MRF) |
+-----------------------------+
^
|
+---------------------------------+
| Model Training & Specialization |
| (MTSF) |
+---------------------------------+
^
|
+-----------------------------------+
| Telemetry Feedback Function (TFF) |
+-----------------------------------+
Intent Reception: User or application submits a high-level intent through the Intent Interface.¶
Adapter Orchestration: IPOF parses the intent, queries the MRF for relevant adapters, and orchestrates dynamic fusion via MFCF.¶
Model Fusion: MFCF composes adapters into a composite model tailored to the intent.¶
Deployment: The composite model is deployed into the IBN system to generate specific configurations.¶
Telemetry Feedback: The TFF monitors the applied configurations' impact on network KPIs and feeds insights back to IPOF and MTSF.¶
Re-Specialization: If performance deviates from expected thresholds, MTSF initiates LoRA re-training or fine-tuning using new data collected via TFF.¶
This architecture closely follows the modularity principles of NWDAF's AnLF for real-time analytics and MTLF for model training and distribution, enabling scalability, continuous learning, and rapid specialization across distributed edge and core network nodes. Each function can be independently scaled and containerized, ensuring fault tolerance and efficient resource utilization. This logical separation allows efficient model lifecycle management, supports federated learning, and can integrate with standard IBN orchestrators via REST APIs or 3GPP-compatible interfaces.¶
Ensuring consistent parameterization and preventing interference when fusing adapters.¶
Preventing malicious adapter injection and ensuring adapters respect security constraints.¶
Managing trust, versioning, and deprecation of adapters.¶
Developing self-optimizing pipelines for adapter generation, validation, and retirement.¶
This document has no IANA actions.¶
TODO¶