💻 computer science

Cross-Version Domain Adaptation and Open-Set Attack Screening for Transferable Wi-Fi Intrusion Detection

This paper proposes a cross-version domain adaptation framework that combines IEEE 802.11 feature harmonization, DANN/CORAL alignment, and confidence-based open-set screening to achieve high-accuracy, real-time Wi-Fi intrusion detection across shifting environments while effectively identifying previously unseen attack families with minimal target-domain labeling.

Original authors: Hadeer Hassan Ibrahim, Rasha M. Ismail, Tamer AbdelKader, Alshaimaa Abo-Alian

Published 2026-08-12

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Hadeer Hassan Ibrahim, Rasha M. Ismail, Tamer AbdelKader, Alshaimaa Abo-Alian

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 a security guard at a massive, bustling train station. Your job is to spot troublemakers. You've spent years memorizing the specific uniforms, walk, and luggage of known troublemakers at this particular station. You know exactly what a "Deauthentication" thief looks like here: they wear a red hat, carry a blue bag, and walk fast. But then, the station management changes the rules. The new station (a different version of the same network) has different lighting, different floor tiles, and the troublemakers have swapped their red hats for green beanies and their blue bags for backpacks. If you try to use your old "red hat" rules here, you'll miss the bad guys completely, or worse, you'll start arresting innocent commuters who happen to be wearing red. This is the core problem of Domain Adaptation in computer security: making a security system that learned in one environment work reliably in a new, slightly different one.

To solve this, security experts use Machine Learning, where computers learn patterns from data, and Open-Set Recognition, which is the ability to say, "I don't know what this is, but it looks suspicious," rather than forcing a guess. The paper you are about to read tackles a very specific version of this train station problem: moving from the AWID2 dataset (an older collection of Wi-Fi attack data) to the AWID3 dataset (a newer, more realistic collection). The researchers wanted to know: Can we teach a computer to recognize Wi-Fi attacks in the new environment without having to re-teach it everything from scratch? They found that by using a clever mix of "adversarial training" (where the computer plays a game of hide-and-seek with itself to ignore irrelevant differences) and a tiny bit of help from the new data, they could build a system that is incredibly accurate, fast, and capable of spotting brand-new types of attacks it has never seen before.

The Wi-Fi Detective's Great Migration

Think of Wi-Fi networks as invisible highways where data cars zoom back and forth. Just like on real highways, bad drivers (hackers) try to cause crashes, steal cargo, or impersonate police cars. To catch them, we use Intrusion Detection Systems (IDS), which are like automated traffic cameras that scream "Stop!" when they see a car doing something weird.

For a long time, these cameras were trained on one specific stretch of highway (the AWID2 dataset). They became experts at spotting the bad cars there. But then, the highway was renovated. The road markings changed, the speed limits shifted, and the bad drivers started driving slightly different cars. This is the AWID3 dataset. When the old cameras were moved to the new highway, they started failing. They either missed the bad guys or started screaming at innocent delivery trucks. This is called Domain Shift, and it's a nightmare for network security because real-world networks are always changing.

The Solution: A Shape-Shifting Detective

The authors of this paper built a new kind of detective system designed to survive this migration. Instead of just memorizing the "red hats" of the old highway, they taught the detective to understand the essence of a bad driver, regardless of the hat color.

1. The Translator and the Mirror
First, they created a "translator" to make sure the old and new highways spoke the same language. They manually mapped the features of the old data to the new data, ensuring that a "frame type" in the old world meant the same thing as a "frame type" in the new world.

Then, they used a technique called DANN (Domain-Adversarial Neural Network). Imagine a game of hide-and-seek. The detective tries to learn what a bad driver looks like, while a "mirror" tries to figure out if the detective is looking at the old highway or the new one. The detective is trained to fool the mirror. If the detective can't tell the difference between the old and new highways, it means the detective has learned the universal signs of a bad driver, not just the specific details of the old road. This is Adversarial Alignment.

2. The Covariance Match
To make sure the detective didn't just get lucky, they added a second rule called CORAL. This is like checking the "vibe" of the traffic. Even if the cars look different, the way they move together (their speed and direction patterns) should be similar. CORAL forces the system to align these movement patterns between the old and new data.

3. The Tiny Hint
Here is the secret sauce: The system didn't just guess. The researchers gave the detective a tiny, tiny hint from the new highway. They showed it just 3,000 labeled packets (1,000 normal, 1,000 deauthentication attacks, and 1,000 evil twin attacks) from the new dataset. This is called Label-Efficient Target Calibration. It's like giving the detective a single photo of a "green beanie" thief from the new station. With this tiny hint, the detective could recalibrate its eyes perfectly.

The Results: A Super-Detective

When they tested this new detective on the new highway (AWID3), the results were staggering.

Accuracy: The system got 99.47% of the known attacks right. That's like catching 99 out of 100 bad guys.
Speed: It was lightning fast. It could process 814,000 packets per second. To put that in perspective, it could check every single car on a massive highway in the blink of an eye, taking only 0.0012 milliseconds per packet.
Cost: In security, missing a bad guy is way worse than accidentally stopping a good guy. The researchers calculated a "Weighted Error Cost" to measure this. Their system had a cost of 517, which is a massive improvement over the old methods that had a cost of over 20,000.

The "Open-Set" Superpower: Spotting the Unknown

The coolest part of this paper is what happens when a completely new type of bad driver shows up—one that wasn't in the training photos at all. In the real world, hackers invent new tricks all the time.

The researchers tested their system against eight new attack families that existed in the new dataset but were totally unknown to the old training data (things like "KRACK," "Kr00k," and "Botnet C&C").

The Result: The system didn't try to guess what these new attacks were. Instead, it said, "I don't know this, but it's definitely not normal!" It flagged anomalies in 504 out of 505 of the new attack files.
The Trade-off: Because the system is so sensitive, it did flag some normal traffic as suspicious too. About 10.1% of the normal packets were flagged as false alarms. But the authors argue this is a fair trade: it's better to have a few extra alerts to check than to miss a brand-new, dangerous attack.

Why This Matters

This paper proves that you don't need to rebuild your security system from scratch every time the network changes. By using adversarial learning to ignore irrelevant differences and tiny hints to recalibrate, you can create a system that travels with you.

The authors also checked to make sure the detective wasn't just relying on spurious correlations by memorizing weird patterns. They used Explainable AI (XAI) to look inside the detective's brain. They found that the system was actually looking at real Wi-Fi clues, like "deauthentication flags" and "signal strength," rather than random noise. This gives us confidence that the system is truly understanding the problem, not just memorizing the answers.

In short, this research shows that we can build Wi-Fi security that is not only smart and fast but also flexible enough to handle the messy, changing reality of the real world, spotting both the old tricks and the brand-new surprises.

Technical Summary: Cross-Version Domain Adaptation and Open-Set Attack Screening for Transferable Wi-Fi Intrusion Detection

1. Problem Statement

Conventional Wi-Fi Intrusion Detection Systems (IDS) often exhibit high reliability on identically distributed datasets but suffer severe performance degradation when deployed in real-world environments where traffic distributions shift. This paper addresses the specific challenge of cross-version transferability in Wi-Fi IDS, focusing on the transition from the AWID2 dataset (source domain) to the AWID3 dataset (target domain).

The core difficulty lies in the significant covariate shift between these datasets, which were collected years apart under different infrastructures, devices, and protocol conditions. Standard supervised learning models trained on AWID2 fail to generalize to AWID3 due to differences in feature distributions, attack semantics, and traffic composition. Furthermore, practical deployment faces two additional hurdles:

Unknown-to-Source Attacks: Target environments may contain attack families (e.g., KRACK, Kr00k, Botnet C&C) that are entirely absent from the source training labels. Treating these as standard classification tasks is unrealistic; instead, they require open-set screening to flag anomalies.
Deployment Constraints: Real-world operations require evaluation beyond accuracy, considering false-positive burdens, inference latency, throughput, and energy consumption, particularly for edge and IoT scenarios.

2. Methodology

The authors propose a unified, transferability-aware framework that integrates semantic feature harmonization, adversarial domain adaptation, and label-efficient calibration. The methodology is structured as follows:

2.1 Semantic Feature Harmonization

To bridge the gap between AWID2 and AWID3, the authors perform a rigorous manual mapping of IEEE 802.11 semantic features.

Feature Mapping: 11 raw feature groups (e.g., frame length, RSSI, sequence numbers, frame control flags) are mapped between datasets.
Derived Indicators: Four protocol-aware boolean flags are generated to explicitly capture management-frame semantics: flag deauth, flag beacon, flag auth, and flag unprotected beacon.
Preprocessing: Continuous features are normalized using source-domain statistics (z-score), while categorical fields undergo one-hot encoding. Missing values are handled with logical defaults to avoid introducing artificial domain identifiers.

2.2 Hybrid Distribution Alignment (DANN + CORAL)

The core architecture utilizes a Domain-Adversarial Neural Network (DANN) to learn domain-invariant representations.

Adversarial Learning: A Gradient Reversal Layer (GRL) is employed to confuse a domain classifier, forcing the feature extractor to ignore domain-specific artifacts (e.g., specific MAC addresses or radio signatures) and focus on structural attack behaviors.
CORAL Alignment: To stabilize training and reduce second-order statistical mismatch, a Correlation Alignment (CORAL) loss is added. This minimizes the distance between the covariance matrices of source and target feature activations.
Dynamic Scheduling: The adversarial weight ( $\lambda$ ) is dynamically scheduled to ensure the model establishes meaningful representations before applying the full domain confusion penalty.

2.3 Label-Efficient Target Calibration

Recognizing that pure zero-shot transfer often fails under extreme domain shifts, the framework employs a label-efficient target calibration stage.

After the main DANN training, a small, balanced subset of labeled AWID3 data (1,000 packets per known class: Normal, Deauthentication, Evil Twin) is used to fine-tune the label predictor.
This step re-anchors the decision boundaries to the target domain without requiring extensive manual labeling.

2.4 Decoupled Inference and Open-Set Screening

The inference mechanism separates known-class classification from unknown-to-source screening:

Known-Class Prediction: Standard softmax classification is applied for known attacks.
Anomaly Screening: Samples with a maximum softmax probability below a high threshold ( $\tau = 0.99$ ) are flagged as unknown-to-source anomalies. This allows the system to detect attack families not present in the AWID2 training set.
MITM Correlation: A post-inference engine correlates sequential events (e.g., deauthentication followed by Evil Twin activity) within the same capture-file session to identify multi-stage Man-in-the-Middle (MITM) attacks.

3. Key Contributions

The paper outlines four primary contributions:

Wi-Fi-Native Domain Adaptation: A framework that mitigates covariate shift between AWID2 and AWID3 through protocol-aware semantic feature alignment, establishing a domain-invariant representation space.
Distribution Alignment with Calibration: A hybrid approach combining adversarial alignment (DANN), covariance matching (CORAL), and a restricted labeled support set for target calibration, demonstrating that label-efficient calibration is critical for performance recovery.
Unknown-to-Source Threat Screening: A confidence-based screening mechanism that successfully flags anomalies in all eight evaluated AWID3-only attack families, moving beyond closed-world classification.
Deployment-Aware Evaluation: A shift from standard accuracy metrics to operational viability, reporting weighted error costs, inference latency, throughput, and class-specific Explainable AI (XAI) analysis to verify reliance on protocol-relevant indicators.

4. Experimental Results

The framework was evaluated on a held-out reduced AWID3 subset and a separate set of unknown-to-source attack files.

4.1 Known-Class Performance

On the held-out AWID3 known-class subset (52,000 packets):

Accuracy: 99.47%
Macro-F1: 0.95
Weighted-F1: 0.99
Recall: 0.99 for Deauthentication and 1.00 for Evil Twin traffic.
Comparison: Baseline models (Random Forest, MLP, 1D-CNN) trained with the same limited target labels achieved significantly lower macro-F1 scores (ranging from 0.26 to 0.57), highlighting the necessity of adversarial alignment.

4.2 Ablation Study

Base DANN (Zero-Shot): Achieved only 54.36% accuracy, confirming that adversarial alignment alone is insufficient for cross-version transfer.
DANN + CORAL: Slight improvement (54.47%), indicating covariance matching helps but does not solve boundary distortion.
Full Model: The addition of label-efficient calibration was the dominant factor, boosting accuracy to 99.47% and macro-F1 to 0.95.

4.3 Unknown-to-Source Screening

File-Level Coverage: Detected anomalies in 504 out of 505 evaluated attack files across eight unknown families (including KRACK, Kr00k, SSH brute force, and Botnet C&C).
Packet-Level Detection: Ranged from 71.2% (Re-association Flood) to 91.4% (SSDP Amplification).
False Positives: On 50,000 held-out known-normal packets, the screening mechanism flagged 10.1% as anomalies. The authors note this is a high-sensitivity configuration suitable for early warning, with practical deployment suggesting temporal aggregation to reduce alert fatigue.

4.4 Operational Metrics

Latency: Average inference latency of 0.0012 ms per packet.
Throughput: Exceeded 814,000 packets per second on the test workstation.
Weighted Error Cost: Reduced from 20,403 (pre-adaptation) to 517 (post-adaptation), calculated with a heavy penalty ( $\alpha=50$ ) for false negatives.
Energy: An illustrative analytical estimate of 0.0031 mJ per packet (based on a 2.5W edge-device budget).

4.5 Explainability

Class-specific Permutation Feature Importance (PFI) analysis confirmed that the adapted model relies on protocol-relevant IEEE 802.11 indicators (e.g., wlan.fc.subtype, flag deauth, RSSI) rather than dataset-specific artifacts.

5. Significance and Claims

The paper claims to shift the Wi-Fi IDS paradigm from closed-world classification toward deployment-aware and transferability-aware intrusion detection.

Operational Viability: The framework demonstrates that high-performance transfer is possible with minimal target-domain labeling, reducing the operational burden of retraining models for new network environments.
Open-Set Awareness: By treating unseen attacks as anomalies rather than misclassified knowns, the system provides early warnings for previously unseen attack families without overstating zero-day classification capabilities.
Protocol-Centric Design: The integration of semantic feature harmonization and XAI ensures that detection decisions are grounded in IEEE 802.11 protocol semantics, enhancing trustworthiness for security practitioners.

The authors modestly acknowledge limitations, noting that the framework relies on manual feature harmonization (which may not scale to future datasets like AWID4 without automation) and that the reported latency/energy metrics are model-level benchmarks rather than full end-to-end embedded device measurements. Future work is proposed to explore automated schema alignment and real-time graph-based correlation for Advanced Persistent Threats (APTs).