Network Exposure Function (NEF) Overview

• NEF is a dual-use concept: in 5G, it securely exposes network capabilities via RESTful APIs, while in causal inference, it maps treatment assignments to exposure states.
• In 5G Service-Based Architecture, NEF mediates API discovery, security, and policy enforcement, translating external requests into 3GPP service invocations.
• In causal inference, NEF formalizes exposure by defining neighborhood treatment patterns, thereby enabling unbiased estimation of peer effects.

Searching arXiv for papers on the two distinct uses of “Network Exposure Function (NEF)”: 5G Network Exposure Function and causal-inference/network-interference exposure functions. Network Exposure Function (NEF) denotes two distinct constructs in current technical literature. In 5G Service-Based Architecture (SBA), the Network Exposure Function is a network function that serves as the single entry point for third-party Application Functions (AFs) to securely access and subscribe to network capabilities through RESTful APIs on the Service-Based Interface (SBI). In causal inference under interference, the Network Exposure Function is a formal rule that maps treatment assignments or peer-treatment patterns into an exposure state, enabling effect estimation when outcomes depend on the treatments of neighboring units rather than on own treatment alone. The two usages are terminologically identical but conceptually separate (Khan et al., 12 Aug 2025, Ugander et al., 2013).

1. Terminological scope

In telecommunications, NEF is part of the 5G Core control plane. Its role is operational and architectural: it authenticates and authorizes AFs, exposes API catalogs, mediates policy, translates RESTful JSON requests into internal 3GPP service invocations, and manages subscriptions for UE events such as mobility or QoS changes. In this sense, NEF is a deployable network function with standardized interfaces and concrete north-bound and south-bound interactions (Khan et al., 12 Aug 2025).

In causal-inference research, NEF is instead a mathematical object. Ugander et al. define exposure sets $\sigma_i^1$ and $\sigma_i^0$ over the global treatment-assignment space and use them to determine whether unit $i$ is “network-exposed to treatment” or “network-exposed to control” (Ugander et al., 2013). A later line of work uses the related term “exposure mapping function” $\phi_e$, defined as a map from peer treatments and graph context to a low-dimensional representation $\rho_i\in\mathbb{R}^d$ for heterogeneous peer-effect estimation (Adhikari et al., 3 Mar 2025).

A common source of ambiguity is therefore nominal rather than substantive. The 5G NEF exposes network capabilities to applications; the causal NEF exposes analytical units to treatment configurations induced by a network.

2. NEF in 5G Service-Based Architecture

Within the 5G SBA, the NEF serves as the single entry point for third-party AFs to securely access and subscribe to network capabilities. It sits on the SBI and exposes RESTful APIs that authenticate and authorize AFs, translate AF subscriptions or queries into 3GPP internal service invocations to functions such as PCF and UDM, enforce policy and charging controls as defined by the Policy Control Function, and interact with the Unified Data Management for user- or subscription-specific data. Its stated core responsibilities are API Discovery and Exposure, Security Enforcement, Policy Mediation, Data Translation, and Event Notification (Khan et al., 12 Aug 2025).

The interaction flow is correspondingly control-plane oriented. An AF requests an OAuth2 token from NEF, NEF retrieves credentials from UDM, and the AF subsequently uses the bearer token to query or subscribe through NEF. A simplified sequence given in the literature includes POST /nef/v1/access-token, credential retrieval from UDM, and GET /nef/v1/subscriptions, with policy validation delegated to PCF before the subscription list is returned to the AF (Khan et al., 12 Aug 2025).

A CAPIF-based proof of concept for vehicular video streaming places NEF on the north-bound interface of the 5G Core, exposed via CAPIF, with south-bound interactions toward PCF, SMF, and indirectly AMF. In that setting, the cited south-bound 3GPP interfaces are Nnef_PolicyAuthorization for PCF, Nnef_PFService for SMF, and Nnef_EventExposure for AMF. The PoC uses four NEF APIs published in CAPIF and discovered by the enhanced VLC client: the PDTQ Policy Negotiation API, the Application Server Session with QoS API, the Event Monitoring API, and the Traffic Influence API (Piscione et al., 3 Aug 2025).

3. API mediation, CAPIF workflows, and LLM-based automation

The CAPIF-mediated PoC describes three textual message-flow sequences. For monitoring network performance, the VLC client performs CAPIF discovery, posts an event subscription, and receives notifications from NEF after PCF or AMF propagates cell-congestion or throughput-related events. For on-demand QoS modification, the client invokes the PDTQ API, NEF forwards the request to PCF as Nnef_PolicyAuthorizationRequest, and the granted decision results in a new QoS profile that SMF applies to the PDU session. For traffic redirection to the edge, the client calls the Traffic Influence API, NEF requests PDU-session modification from SMF, and UPF forwarding rules are updated so that the path shifts from a core PSA UPF to a local edge UPF, reducing latency (Piscione et al., 3 Aug 2025).

The same API-exposure setting motivates automation work. NEFMind uses 3GPP-standard NEF API OpenAPI YAML files as input artifacts and constructs a synthetic fine-tuning dataset through a seven-step process: flattening YAML so each endpoint definition is self-contained; prompting GPT-4 for Request–Response JSON pairs; generating initial JSON objects covering seven NEF endpoints; manually validating and pruning fabricated entries; scaling each request string to 100 paraphrases for approximately 765 total records; splitting the data into train and eval partitions of approximately 70% and 30%; and converting JSON to CSV and Instruct–Output format for Phi-2 fine-tuning (Khan et al., 12 Aug 2025).

Model adaptation is performed with Quantized-Low-Rank Adaptation (QLoRA), where pretrained weights remain 4-bit quantized and only a low-rank update is learned:

$\Delta W = B\cdot A,$

with $A\in \mathbb{R}^{r\times d}$, $B\in \mathbb{R}^{d\times r}$, and $r\ll d$. The reported training objective is

$$\mathcal{L}(\theta + \Delta W) = \sum_{i=1}^N \ell\bigl(f_{\theta + BA}(x_i),\,y_i\bigr) + \lambda\|B\|_F^2 + \lambda\|A\|_F^2,$$

where $\sigma_i^0$0 is cross-entropy loss over token sequences. The configuration highlights are LoRA $\sigma_i^0$1, dropout $\sigma_i^0$2, rank $\sigma_i^0$3, target modules q_proj, k_proj, v_proj, dense, fc1, and fc2, with task type causal LM (Khan et al., 12 Aug 2025).

Performance is reported with GPT-4 Ref Score, BertScore, and communication-overhead reduction. Baseline Phi-2 attains GPT-4 Ref scores from 4.78 to 10.43 and BertScore from 0.7657 to 0.7699, whereas the fine-tuned Phi2-NEF attains GPT-4 Ref scores from 98.26 to 100.00 and BertScore from 0.9979 to 0.9986. Manual discovery is treated as a 100% signaling baseline, while NEFMind automated discovery uses 15% of baseline signaling, corresponding to an 85% reduction. The open-source Phi-2 model is reported to achieve 98–100% API call identification accuracy and performance comparable to GPT-4 while remaining computationally efficient for telecommunications infrastructure deployment (Khan et al., 12 Aug 2025).

Deployment details are equally explicit. Fine-tuning is reported on an Amazon g5.4xlarge instance with 1 A10G GPU, 24 GB VRAM, 16 vCPU, and 64 GB RAM. The run takes approximately 595 s at 4.5 samples/s and 1.5 steps/s, consumes $\sigma_i^0$4 FLOPs, and peaks at approximately 22 GB VRAM. Inference latency for Phi-2 on A10G is approximately 150–200 ms per request, versus approximately 800–1200 ms plus queuing for the GPT-4 public API. For approximately 1000 QPS, the reported guideline is to provision 4–8 A10G GPUs with load balancing and to use batch sizes 8–16 when real-time constraints allow (Khan et al., 12 Aug 2025).

4. NEF in causal inference under interference

In graph-based experimentation, let $\sigma_i^0$5 be an interference graph on $\sigma_i^0$6 vertices and let $\sigma_i^0$7 denote the random treatment-assignment vector. For each vertex $\sigma_i^0$8, Ugander et al. define exposure sets $\sigma_i^0$9 and $i$0, with $i$1 and $i$2. The Network Exposure Indicator is then

$i$3

and is undefined otherwise, or treated as neither exposure. This formulation makes the exposure condition explicit at the level of global assignments rather than individual treatment alone (Ugander et al., 2013).

Two canonical exposure models are given. Under $i$4-hop full neighborhood exposure, with $i$5, treatment exposure requires $i$6 for all $i$7, and control exposure is defined analogously with zeros. Under fractional $i$8-neighborhood exposure, with $i$9 and $\phi_e$0, treatment exposure is

$\phi_e$1

with the corresponding control condition requiring $\phi_e$2 and at least $\phi_e$3 untreated neighbors. These definitions operationalize interference assumptions as graph-theoretic predicates over assignments (Ugander et al., 2013).

The same work studies graph-cluster randomization. Vertices are partitioned into clusters $\phi_e$4, each cluster assignment $\phi_e$5 is drawn independently as $\phi_e$6, and then all vertices in cluster $\phi_e$7 inherit $\phi_e$8. For fractional-$\phi_e$9 exposure, if $\rho_i\in\mathbb{R}^d$0 indexes the clusters intersecting $\rho_i\in\mathbb{R}^d$1, if $\rho_i\in\mathbb{R}^d$2, and if $\rho_i\in\mathbb{R}^d$3 denotes the cluster containing $\rho_i\in\mathbb{R}^d$4, the exposure probability $\rho_i\in\mathbb{R}^d$5 is the probability of a weighted Poisson-binomial event. The paper gives an $\rho_i\in\mathbb{R}^d$6 recursion for computing it exactly:

$\rho_i\in\mathbb{R}^d$7

This is used to obtain the exposure probabilities needed for inverse-probability weighting (Ugander et al., 2013).

With $\rho_i\in\mathbb{R}^d$8, Ugander et al. use a Horvitz–Thompson estimator for the average treatment effect

$\rho_i\in\mathbb{R}^d$9

namely

$\Delta W = B\cdot A,$0

The estimator is unbiased provided the exposure model correctly recovers the relevant potential outcomes. The variance analysis then shows a sharp contrast between designs: under independent-vertex randomization and full 1-hop exposure, $\Delta W = B\cdot A,$1, and the variance can grow exponentially in degree; on restricted-growth graphs, a 3-net clustering yields a bound

$\Delta W = B\cdot A,$2

where the constants depend on $\Delta W = B\cdot A,$3 but not on $\Delta W = B\cdot A,$4 or $\Delta W = B\cdot A,$5 (Ugander et al., 2013).

5. Learned exposure mapping functions

A more recent formulation treats the Network Exposure Function as a learned representation rather than a hand-specified indicator. On an undirected graph $\Delta W = B\cdot A,$6 with node covariates $\Delta W = B\cdot A,$7, edge attributes $\Delta W = B\cdot A,$8, binary treatment vector $\Delta W = B\cdot A,$9, and outcomes $A\in \mathbb{R}^{r\times d}$0, the exposure mapping function $A\in \mathbb{R}^{r\times d}$1 is defined by

$A\in \mathbb{R}^{r\times d}$2

where $A\in \mathbb{R}^{r\times d}$3 is the neighborhood of unit $A\in \mathbb{R}^{r\times d}$4. Under neighborhood interference, $A\in \mathbb{R}^{r\times d}$5 compresses the high-dimensional pattern of treated peers, together with local graph context, into a $A\in \mathbb{R}^{r\times d}$6-dimensional peer-exposure representation. In the reported implementation, $A\in \mathbb{R}^{r\times d}$7 (Adhikari et al., 3 Mar 2025).

The proposed EgoNetGnn architecture learns $A\in \mathbb{R}^{r\times d}$8 with two GNN-based modules run in parallel. One feature-embedding GNN processes node covariates and edge covariates to produce a learned confounder or effect-modifier representation $A\in \mathbb{R}^{r\times d}$9. In parallel, an ego-network-based exposure-mapping GNN is constructed for each focal unit by extracting the ego graph over its neighbors, initializing peer-node states from treatment and edge-to-ego attributes, performing unlabeled neighborhood aggregation inside the ego graph, and then passing the aggregated states through an MLP encoder before pooling. The paper states that this design enables the model to count or otherwise encode small closed-triad motifs, component counts, and tie-strength patterns by turning triangles into simple edges in the ego graph (Adhikari et al., 3 Mar 2025).

Training is end-to-end with a two-head TARNet estimator for the counterfactual outcome surfaces $B\in \mathbb{R}^{d\times r}$0 and $B\in \mathbb{R}^{d\times r}$1. The per-node loss is

$B\in \mathbb{R}^{d\times r}$2

where $B\in \mathbb{R}^{d\times r}$3 is the factual MSE, the balance loss encourages the learned $B\in \mathbb{R}^{d\times r}$4 across the batch to look uniform in $B\in \mathbb{R}^{d\times r}$5, and the $B\in \mathbb{R}^{d\times r}$6 term promotes invariance to irrelevant attributes. The reported hyperparameters are $B\in \mathbb{R}^{d\times r}$7 and $B\in \mathbb{R}^{d\times r}$8, selected by held-out validation with a 20% split (Adhikari et al., 3 Mar 2025).

The downstream target is heterogeneous peer-effect estimation. The individual peer effect is defined as

$B\in \mathbb{R}^{d\times r}$9

and the model estimates the corresponding conditional means through the TARNet heads after substituting the learned $r\ll d$0. Accuracy is evaluated with Precision in the Estimation of Heterogeneous Effects,

$r\ll d$1

On Watts–Strogatz, Barabási–Albert, and Stochastic-Block graphs under mutual-connection exposure, baselines based on simple treated fraction or attention weights incur log-scale $r\ll d$2 10–100× larger than EgoNetGnn. The method also significantly outperforms baselines under connected-component exposure, matches or exceeds motif-based alternatives under clustering-coefficient exposure, and achieves the lowest $r\ll d$3 in semi-synthetic BlogCatalog experiments across all tested exposure mechanisms (Adhikari et al., 3 Mar 2025).

6. Conceptual relations, misconceptions, and future directions

The two NEF traditions should not be conflated. In 5G systems, NEF is a standardized exposure and mediation layer for network capabilities, policy control, and event notification. In causal inference, NEF is a formalization of how treatment assignments on a graph induce the exposure state relevant to an outcome. The shared phrase “network exposure” is therefore accidental at the acronym level rather than evidence of a common formalism.

At the same time, both literatures are concerned with controlled reduction of high-dimensional network context. In telecommunications, NEF translates heterogeneous internal service invocations and policy logic into north-bound APIs consumable by AFs, and recent work further compresses NEF API ecosystems into synthetic training corpora suitable for parameter-efficient LLM adaptation (Khan et al., 12 Aug 2025). In causal inference, NEF or $r\ll d$4 reduces a combinatorially large treatment-neighborhood configuration into an analytically usable exposure condition or embedding (Ugander et al., 2013, Adhikari et al., 3 Mar 2025). This suggests a structural analogy: both senses of NEF define an interface between complex network state and downstream decision procedures, although the objects being exposed are entirely different.

Research directions are most explicit on the telecommunications side. Reported future avenues include scaling synthetic dataset size with active learning to avoid overfitting, integrating formal verification layers to guarantee API schema compliance, extending the approach to other 5G Core NFs such as NRF, AUSF, and SMF, incorporating continuous-learning pipelines for API specification evolution, and adding adversarial hardening against malformed or malicious AF requests (Khan et al., 12 Aug 2025). A plausible implication of the causal-inference literature is that exposure modeling is moving from fixed, hand-specified treated-fraction summaries toward learned representations that incorporate local neighborhood structure and edge attributes when peer influence mechanisms are unknown (Adhikari et al., 3 Mar 2025).