💻 computer science

Explanation-Surface Governance in XAI-Enabled Network Intrusion Detection under Resource Constraints

This paper introduces XAI-Gate, a novel, detector-agnostic governance framework that dynamically selects explanation actions to balance packet quality of service, operator interpretability, and information exposure under resource constraints, achieving significant reductions in explanation debt and zero budget violations without requiring classifier re-training.

Original authors: Tran Duc Le, Truong Duy Dinh, Phuc Hao Do, Van Dai Pham

Published 2026-07-30

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Tran Duc Le, Truong Duy Dinh, Phuc Hao Do, Van Dai Pham

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the captain of a massive, high-speed spaceship flying through a galaxy full of invisible space pirates. Your ship has a super-smart computer that can spot these pirates instantly. But here's the catch: the computer is so fast it spots thousands of potential threats every second. If the captain tried to read a full, detailed report for every single "maybe-pirate," the ship's engines would stall, the crew would get overwhelmed, and the ship might crash. This is the world of Network Intrusion Detection. It's the digital equivalent of that spaceship, constantly scanning for hackers trying to break into our computers and phones.

To make sense of all these alerts, scientists use something called Explainable AI (XAI). Think of XAI as a translator that turns the computer's scary, confusing "I think this is bad" into a human-readable note like, "This looks like a pirate because they are wearing a red hat and moving too fast." This helps security guards (the operators) decide what to do. But there's a problem: writing these notes takes time and energy. If the computer tries to write a perfect, detailed note for every single alert, it runs out of battery, gets slow, and might even accidentally reveal its own secrets to the pirates. The big question this paper tackles is: How do we manage the flow of these notes so the ship keeps moving, the crew stays sane, and the pirates don't learn our secrets?

The Problem: Too Many Notes, Not Enough Time

In the real world, security systems are often under attack. A clever hacker doesn't just try to break in; they can flood the system with thousands of "borderline" alerts—tricky situations that aren't clearly good or bad. This is like a pirate throwing thousands of fake red hats at the ship's sensors. If the security computer tries to explain every single one of these fake alerts, it gets bogged down. It creates a backlog of "explanation debt" (a pile of unexplained alerts), slows down the actual traffic (the ship's engines), and might accidentally spill too much information about how the defense works.

Most current research focuses on making the explanations better or smarter. But this paper argues that making them smarter isn't enough. You also need a manager to decide which explanations to write, when to write them, and how much detail to include.

The Solution: XAI-Gate, the Traffic Cop

The authors introduce a new system called XAI-Gate. Imagine XAI-Gate as a super-smart traffic cop standing between the security computer and the human operators. Instead of letting every alert through to get a full explanation, XAI-Gate looks at the current situation and chooses one of seven actions for each alert:

Suppress: "Ignore this one; it's probably nothing."
Coarse: "Here's a quick summary: 'Red hat detected,' but no details."
Delay: "Hold on, we're too busy right now. I'll explain this later."
Full Local: "Here is the full, detailed report right now."
Offload: "This is too heavy for us; let's send it to a cloud server to handle."
Audit-Only: "Write it down in a secret log for later, but don't show it to anyone yet."
Redact: "This is suspicious, but I'm going to hide the sensitive parts so we don't accidentally tell the hacker how we think."

XAI-Gate doesn't try to be a better detective. It doesn't even try to be a better translator. It just manages the flow. It looks at how busy the system is, how much "debt" (unexplained alerts) has piled up, and how much "exposure budget" (how much secret info we can afford to leak) is left. If the system is swamped, XAI-Gate might say, "Okay, let's just give a quick summary for this one and redact the details for that one," to keep the ship moving.

What They Found: A Trade-Off, Not a Magic Wand

The researchers tested XAI-Gate using a rigorous stress-test called XAI-SurfaceBench. They simulated different types of traffic, including normal days, busy days, and days where hackers were actively trying to overwhelm the system with fake alerts.

Here is what the simulations showed:

The "Always Explain" Trap: If you try to explain everything (the "Always Explain" policy), you get perfect coverage (100% of alerts explained), but you blow your budget. In their tests, this approach used 2.864 times the allowed exposure budget and violated safety limits in 65.1% of the time slots. It's like trying to read every single book in a library at once; you get the info, but you burn out the lights and the building collapses.
The "Never Explain" Trap: If you explain nothing, you save energy, but you have a massive pile of "explanation debt" (over 20,000 units in their test), leaving the operators in the dark about real threats.
The XAI-Gate Sweet Spot: XAI-Gate found a middle ground. In the default "adversarial" test (where hackers were trying to inflate the demand), XAI-Gate managed to explain 0.800 (or 80%) of the high-risk alerts. Crucially, it did this while keeping the exposure budget exactly where it should be (zero violations) and cutting the explanation debt by 57.3% compared to standard methods.

However, the paper is very clear about what XAI-Gate is not. It is not a magic shield that stops all attacks. When the hackers pushed the demand to extreme levels (a multiplier of 2.0), even XAI-Gate couldn't save everything. The high-risk coverage dropped to 0.223, and packet drops (lost data) appeared. The authors call this a "Denial-of-Explanation" state. It's a hard limit: if the enemy throws enough fake hats at the ship, no amount of traffic management can explain them all without crashing the system.

The Bottom Line

This paper suggests that we need to stop treating explanations as free gifts that we can just hand out. Instead, we need to treat them as a limited resource, like fuel or time. XAI-Gate is a new way to govern that resource. It proves that by being smart about which explanations to give and how much detail to share, we can keep our digital ships running smoothly, keep our secrets safe, and still give our operators the information they need to catch the bad guys. It's not about finding a better detective; it's about hiring a better traffic cop.

Technical Summary: Explanation-Surface Governance in XAI-Enabled Network Intrusion Detection under Resource Constraints

1. Problem Statement

Current research in Explainable AI (XAI) for Network Intrusion Detection Systems (NIDS) predominantly focuses on semantic interpretability—making model decisions transparent to analysts. However, this literature largely treats explanations as passive trust artifacts appended to detection outputs, ignoring their operational reality as resource-intensive service outputs. In resource-constrained environments (e.g., edge, IoT, and enterprise gateways), generating explanations consumes CPU, memory, and accelerator resources, introduces queueing delays, and creates "explanation debt" (a backlog of unexplained or low-fidelity alerts).

Furthermore, the explanation interface itself presents a security vulnerability. Adversaries can manipulate traffic to trigger "explanation-demand inflation," generating a high volume of borderline alerts that force the system to consume excessive compute, delay critical alerts, or exhaust information-exposure budgets. Existing frameworks lack a mechanism to govern when, how, and at what fidelity an explanation should be generated when facing finite capacity, bursty traffic, and adversarial pressure. The core problem is the absence of a control layer that balances packet Quality of Service (QoS), operator interpretability, explanation debt, and information leakage exposure.

2. Methodology: XAI-Gate

The paper introduces XAI-Gate, a detector-agnostic online service-governance framework that sits between the intrusion detector and the explanation delivery mechanism. It does not require retraining the underlying IDS or the explainer. Instead, it manages the "explanation surface" by dynamically selecting from a set of seven explanation actions based on real-time system state.

2.1 System Model and State

The system is modeled as a constrained stochastic control problem (Markov Decision Process). The state $S_t$ at time slot $t$ includes:

Packet Queue Pressure ( $Q_t$ ): Current load on the packet service queue.
Explanation Debt ( $E_t$ ): A risk-weighted backlog of alerts requiring explanation.
Exposure Budget ( $L_t$ ): The remaining allowable information leakage.
Offload Belief ( $b_{rt}$ ): Estimated Round-Trip Time (RTT) for offloading to fog/cloud.
Suspicion State ( $H_t$ ): A score indicating potential adversarial explanation-demand inflation.

2.2 Action Space

For each alert, XAI-Gate selects one action from the set $\mathcal{A} = \{0, C, D, F, O, A, R\}$ :

Suppress (0): No explanation; avoid low-value or unsafe releases.
Coarse (C): Disclose top feature groups only; limits detail under medium suspicion.
Delay (D): Hold explanation in a queue until resources recover.
Full Local (F): Release full local attribution (e.g., SHAP/LIME) for maximum detail.
Offload (O): Send sanitized alert data to a managed cloud/fog explainer.
Audit-Only (A): Log metadata for forensics without immediate release.
Redact (R): Mask sensitive attributes under high suspicion or exposure pressure.

2.3 Decision Policy

The policy $\pi$ maps the system state and alert features (score $s$ , estimated cost $c$ ) to an action. The selection is driven by an action score $\Psi_t(a; i)$ , which balances:

Utility: Security value and operator usefulness of the explanation.
Costs: Local compute cost, exposure penalty, debt increment, and offload RTT.
Constraints: Hard caps on cumulative exposure budget and local compute capacity.

The policy includes a "hard exposure-cap guard" that downgrades or delays explanations if admitting a full set of high-risk alerts would overshoot the remaining exposure budget. The method is implemented via an online surrogate inspired by constrained MDP and drift-plus-penalty control, ensuring stability without requiring a known transition law.

3. Key Contributions

Explanation-Surface Management Formulation: The paper reframes XAI-enabled defense as a service-management problem where explanations are finite-capacity, delay-sensitive, and adversarially triggerable outputs. It defines the managed state, action space, and threat models for explanation-demand inflation.
XAI-Gate Online Control Policy: A detector-agnostic policy that dynamically selects explanation fidelity and delivery mode based on packet pressure, debt, offload uncertainty, and suspicion. It enforces hard exposure caps without requiring new IDS training.
Formal Service-Boundary Analysis: The authors derive dominance conditions and feasibility propositions (e.g., when full local explanations are dominated by cheaper actions, or when offloading becomes preferable) that make the control logic inspectable rather than purely heuristic.
XAI-SurfaceBench: A rigorous stress-test benchmark anchored on KDDCup99, UNSW-NB15, and TON IoT datasets. It evaluates performance across nine traffic/uncertainty regimes (including bursty, non-Markovian, and adversarial demand inflation) without requiring new detector training, focusing instead on service-frontier behavior.

4. Experimental Results

The evaluation was conducted using 20 random seeds across various stress regimes. Key findings include:

Adversarial Demand Inflation: In the default adversarial regime, XAI-Gate achieved a high-risk explanation coverage of 0.800 while maintaining zero exposure-budget violations. In contrast, conventional threshold policies achieved slightly higher coverage (~0.884) but violated the exposure budget in 28% of slots and consumed 39% more exposure budget.
Debt Reduction: Compared to threshold policies, XAI-Gate reduced mean explanation debt by 57.3% (from ~3660 to ~1561).
Service Frontier: XAI-Gate does not universally maximize coverage. Instead, it shifts the operating frontier, trading a small amount of immediate high-risk coverage (approx. 8.4 percentage points) for bounded exposure and significantly lower debt.
Robustness: Under severe demand inflation (attack multiplier 2.0), the system enters a "Denial-of-Explanation" (DoE) state where coverage drops and packet drops occur, confirming that finite capacity is a physical limit that no policy can bypass. However, XAI-Gate consistently maintained zero exposure violations in feasible regimes where baselines failed.
Ablation Studies: Removing specific mechanisms (debt awareness, exposure budget, or fidelity control) degraded performance, confirming that the joint management of these variables is essential for the observed gains.
Real-Data Anchors: Results remained consistent when calibrated using UNSW-NB15 (including a de-duplicated rerun) and TON IoT datasets, reinforcing that the findings are not artifacts of synthetic score distributions.

5. Significance and Claims

The paper modestly positions its contribution not as a new detector, a new explainer, or a universal accuracy improvement, but as a service-management layer for XAI-enabled network defense.

Operational Shift: The primary significance is the shift from treating explanations as static interpretability artifacts to treating them as dynamic, resource-constrained service outputs that require admission control and fidelity selection.
Bounded Feasibility: The work demonstrates that it is possible to govern explanation generation to prevent information leakage and service degradation without retraining the underlying AI models.
Trade-off Management: The findings show that in constrained environments, the optimal strategy is not "always explain" or "explain based on a static threshold," but a state-aware policy that balances packet QoS, operator context, and security exposure.
Limitations: The authors explicitly state that XAI-Gate cannot improve the baseline detection accuracy (ROC) of the underlying IDS. Furthermore, the evaluation relies on simulated queueing and proxy timing; absolute hardware latency and human operator trust dynamics require further validation through hardware-in-the-loop and user studies. The system is designed for constrained settings; in environments with abundant resources or no exposure concerns, simpler policies may suffice.

In summary, XAI-Gate provides a reproducible framework for controlling the "explanation surface," ensuring that the operational benefits of XAI in NIDS do not come at the cost of system stability, resource exhaustion, or security exposure.