WiFireNet Handgun Detection Works on Old Wi-Fi, but Its Best Result Hides the Hard Part

WiFireNet handgun detection reached 72.5% average precision using signals based on the 1999-era 802.11a standard. That result came from a controlled indoor experiment with no more than four moving people. It is an intriguing research result, but it is not evidence that ordinary routers can reliably screen public spaces.

Researchers at Sweden’s Defence Research Agency, known as FOI, introduced WiFireNet in a paper submitted on October 13, 2026. Their model reads channel state information, or CSI, which describes how a wireless channel changes a transmitted signal. The system tries to identify people carrying concealed handguns by learning the radio patterns produced by bodies and carried objects.

The project challenges the usual choice between cameras and checkpoint scanners. Cameras cannot see an object hidden under clothing, while metal detectors constrain people to designated screening lanes. WiFireNet explores a third route: turning radio signals already moving through a room into a complementary sensor.

The reversal is that the older, simpler Wi-Fi configuration produced the strongest handgun result. An expanded experimental configuration collected more direct channel measurements, yet performed worse across the broader detection task. That finding makes the work more interesting, but it also exposes how strongly performance depends on the room, dataset, and evaluation design.

WiFireNet Handgun Detection Turns Radio Measurements Into Visual Predictions

FOI’s experiment does not create a picture through Wi-Fi. It converts radio measurements into a representation that a vision-style detector can process.

The researchers built the WiFire dataset by recording wireless signals and synchronized RGB video while participants moved through indoor rooms. The video supplied labels during training and evaluation. At inference, the final WiFireNet model operated only on CSI samples, according to the WiFireNet paper.

CSI contains amplitude and phase measurements across portions of a wireless channel. Amplitude describes signal strength, while phase records a wave’s position within its cycle. Together, those measurements reflect absorption, scattering, and multipath propagation inside a room.

Multipath propagation occurs when a signal reaches a receiver through several reflected routes. A person, wall, bottle, phone, or handgun can alter those routes. The model searches for repeatable differences within the resulting measurements.

The team collected four carried-object classes: a deactivated metal handgun, a plastic handgun, a metal water bottle, and a mobile phone. Those choices create deliberate confusion. The plastic replica shares the handgun’s general shape, while the bottle and phone introduce familiar materials and carried-object profiles.

At most one person in each recorded scene carried the deactivated metal handgun. The larger collection area measured 12 square meters and contained zero to four people. A second room measured eight square meters and contained no more than two people.

The researchers used two laptops, with one connected to transmitters and another connected to receivers. A USB camera attached to the receiving computer captured video for annotation. CSI arrived in bursts at 80 Hz, while the camera recorded at 30 frames per second.

The main WiFi-regular and WiFi-plus configurations each contained about 100,000 CSI samples. The small-room collection contained about 40,000 samples. These are individual radio observations, not 240,000 independent rooms, participants, or deployment trials.

WiFireNet processes the three latest CSI samples for each prediction. The authors describe that input as covering 37.5 milliseconds. The model separates phase from amplitude and sends each component through its own convolutional encoder.

A modality translation network then reshapes the combined features into a synthetic spatial representation. This representation grows through several upsampling stages until it matches the camera image dimensions. It is a learned coordinate surface, not a photograph reconstructed from radio waves.

A feature pyramid network examines that surface at multiple spatial scales. A region proposal network suggests possible bounding boxes, and a detection head refines and classifies them. These parts borrow heavily from established computer-vision object detectors.

The camera therefore plays an important role even though it is absent during final inference. A vision model acts as a teacher during training through cross-modal knowledge distillation. That process encourages Wi-Fi features to resemble useful features extracted from synchronized images.

FOI has published the project’s open-source code and released the WiFire dataset. That gives other researchers a chance to inspect the design, rerun experiments, and test more demanding data splits.

The immediate change is not a deployable security product. It is a reproducible research package that frames concealed-handgun detection as a dynamic, multiperson Wi-Fi sensing problem.

The 72.5% Result Is Average Precision, Not a Security Success Rate

The headline number measures ranked detection performance under one evaluation condition. It does not mean the system correctly clears or flags 72.5% of people.

For the WiFi-regular configuration, handgun-class average precision reached 72.5% at a 25% intersection-over-union threshold. At a stricter 50% threshold, the handgun score fell to 53.8%.

Intersection over union, or IoU, measures how much a predicted box overlaps the labeled box. A 25% threshold accepts relatively loose localization. A 50% threshold demands a closer match between the prediction and its reference.

Average precision combines precision and recall across confidence thresholds. It rewards a model that ranks correct detections above false ones. It is not interchangeable with accuracy, and it does not directly describe a checkpoint’s false-alarm rate.

That distinction matters because a deployed security system operates at a selected alert threshold. Operators need to know how many armed people it misses and how many unarmed people it flags. The paper reports research metrics, not an operational threshold chosen for public screening.

The 72.5% figure also covers the handgun class alone. When all classes were included, WiFi-regular produced mean average precision of 50.3% at 25% IoU. That result declined to 47.1% at 50% IoU.

Recall for the same configuration was higher. WiFi-regular reported 91.5% recall at 25% IoU and 89.5% at 50% IoU across the evaluated classes. High recall can coexist with substantial false positives, especially when confidence thresholds are adjusted to capture more possible targets.

The stricter handgun result is particularly revealing. WiFireNet retained evidence that a handgun carrier was present, but its predicted location often lacked a close match. The system associates a detection with an image-space region around a person rather than identifying the weapon’s exact position.

That limitation does not erase the underlying result. Detecting a concealed metal handgun among phones, bottles, and a plastic replica is more specific than detecting generic metal. The researchers also moved beyond static, isolated-object experiments by recording multiple people in motion.

However, the experiment still occurred inside two small, controlled rooms. The larger scene covered 12 square meters, an area smaller than many offices, classrooms, or retail entrances. Its maximum crowd contained four people.

Real screening environments create harder class imbalances. Most people will not carry a handgun, so even a modest false-positive rate can generate many alerts. The published average-precision values do not answer how WiFireNet behaves under that operational imbalance.

Security buyers would also need results broken down by clothing, body position, movement direction, handgun orientation, and distance. The current paper does not establish consistent performance across those variables. It offers a baseline for further testing.

The careful conclusion is that WiFireNet handgun detection found usable statistical structure in CSI. The experiment does not establish a dependable probability that a person is armed.

That difference separates an encouraging machine-learning benchmark from a security decision. A model can perform well enough to justify more research while remaining unsafe for autonomous deployment.

More Wi-Fi Measurements Produced Worse Results

WiFireNet’s most instructive result is not its highest score. It is the failure of additional direct channel measurements to improve detection.

The team collected data under two radio configurations. WiFi-regular followed the standard IEEE 802.11a arrangement and used four pilot subcarriers. Values for the remaining 48 subcarriers were obtained through linear interpolation.

WiFi-plus instead used 48 of the 52 available subcarriers as pilots. That provided direct measurements across 92.3% of the available subcarriers. The configuration was intended to offer denser channel information and move closer to radar-style sensing.

A simple expectation follows: richer measurements should make objects easier to distinguish. The results did not support that expectation. WiFi-plus reached only 39.9% mean average precision at 25% IoU and 37.2% at 50% IoU.

WiFi-regular scored 50.3% and 47.1% under those same thresholds. Its handgun-only result also led by a wide margin. WiFi-plus reported 44.2% handgun average precision at both IoU thresholds, compared with WiFi-regular’s 72.5% and 53.8%.

The authors do not claim a definitive explanation. They suggest that the relatively clean collection environment generated fewer useful multipath reflections. Extra subcarrier data might matter more inside a complex, cluttered room.

Random differences between datasets are another possibility. WiFi-regular and WiFi-plus were collected as separate configurations, so their comparison does not isolate radio resolution from every other source of variation. The paper identifies the result as a question for future work.

This ambiguity illustrates a central challenge in Wi-Fi sensing. More signal detail does not automatically produce more transferable information. A learning system can exploit patterns tied to a room, collection session, or hardware path rather than the intended object.

The small-room results reinforce that problem. Its baseline mean average precision reached 49.8% at 25% IoU and 46.9% at 50% IoU. Those figures closely approached WiFi-regular despite using the WiFi-plus configuration.

The authors attribute easier localization in the smaller room to stronger multipath reflections from nearby walls. In other words, the room can improve sensing by creating a richer set of radio paths.

That mechanism also makes deployment difficult. Moving furniture, opening a door, adding partitions, or changing crowd density can reshape those paths. A model trained in one layout can encounter a different input distribution without leaving the building.

The network design itself produced clearer gains. Replacing the convolutional modality translator with a parameter-matched multilayer perceptron reduced WiFi-regular mean average precision from 50.3% to 35.5% at 25% IoU.

At 50% IoU, the same change reduced performance from 47.1% to 31.3%. Convolution appears to help the network capture local structure within the phase and amplitude inputs.

Vision-based knowledge distillation also helped some configurations. Removing that loss reduced WiFi-regular performance from 50.3% to 42.4% at 25% IoU. At 50% IoU, the result declined from 47.1% to 39.4%.

However, the distillation loss did not improve WiFi-plus. That uneven contribution again suggests a system whose behavior depends on the collection condition.

The main mechanism is therefore credible but unsettled. CSI contains information related to people and carried objects. WiFireNet can translate some of that structure into spatial detections.

What remains unknown is whether the learned representation describes handguns broadly or memorizes stable correlations within these rooms. More subcarriers did not resolve that uncertainty.

The Real Opponent Is Room Dependence, Not Metal Detectors

WiFireNet is best understood as a complementary sensing experiment whose primary obstacle is environmental generalization.

The paper positions Wi-Fi sensing beside cameras and metal detectors. That framing is reasonable because each modality sees different evidence. Cameras capture appearance, metal detectors react to conductive material, and CSI reflects changes across a radio channel.

Yet WiFireNet does not currently compete with mature screening hardware on reliability. Its more immediate contest is between room-specific signal learning and a detector that generalizes across locations.

The training setup makes that contest especially important. Researchers split the datasets at the level of individual CSI samples. Half of the samples went to training, while the other half formed the test set. Twenty percent of the training portion served as validation data.

Individual samples recorded at 80 Hz can be closely related in time. Neighboring observations may show the same people, objects, positions, and room state. A random sample-level split can therefore place highly similar moments on opposite sides of the training boundary.

The paper follows precedents from related Wi-Fi sensing research, but that does not settle the deployment question. A stronger evaluation would hold out complete participants, recording sessions, rooms, or hardware arrangements.

The study does not report a room-held-out evaluation. It also does not describe testing the trained model inside an unfamiliar building. Consequently, claims about performance in unseen layouts remain unsupported.

Tom’s Hardware noted that the prototype struggles in unfamiliar indoor configurations. That limitation follows directly from the role of multipath structure. The radio environment is not background noise for WiFireNet; it is part of the model’s input.

Earlier work shows that the broader idea is not new. A 2018 commodity Wi-Fi study explored non-obtrusive detection of concealed metallic objects. Other projects have used CSI for human activity recognition, localization, pose estimation, and people counting.

A peer-reviewed concealed-object study published in 2025 reported 87.5% accuracy for a real gun in indoor testing. That project classified body-plus-weapon cases rather than producing WiFireNet’s multiperson image-coordinate detections.

Those numbers should not be ranked directly. Accuracy and average precision measure different behavior, while the datasets and scene designs also differ. Their coexistence shows that radio-based object sensing has moved beyond a single laboratory demonstration.

WiFireNet advances the line of research through dynamic groups, confusing objects, spatial detections, and public artifacts. Its open dataset also makes independent scrutiny easier than a closed prototype would.

Still, public code cannot compensate for narrow training data. An open model can be inspected and extended, but it inherits the conditions represented in its recordings.

The phrase “ordinary Wi-Fi” also needs qualification. The experiment used a modified open-source IEEE 802.11a implementation in GNU Radio. It operated at 5.68 GHz with 20 MHz of bandwidth.

That setup uses an established wireless standard, but it is not equivalent to installing an application on any consumer router. Deployments would need compatible CSI access, known transmitter and receiver placement, calibrated timing, and software capable of processing the measurements.

Modern Wi-Fi generations can expose wider channels and more detailed measurements. Those capabilities provide research opportunities, but WiFireNet’s own WiFi-plus result warns against assuming that more data guarantees better detection.

The system pressures researchers to solve generalization before discussing broad adoption. It does not yet pressure schools, airports, or offices to replace their existing security equipment.

Open Source Helps Validation, but It Also Sharpens the Privacy Question

A system that senses concealed objects without a camera can reduce one privacy risk while creating another form of invisible observation.

WiFireNet does not produce a recognizable photograph during inference. That property can limit collection of facial appearance, clothing detail, and other visual information. It also supports sensing in darkness or through visual occlusion.

However, privacy does not depend only on whether a camera is present. A radio system can still infer the presence, movement, position, or possessions of people who never interact with it.

Wi-Fi sensing can be difficult to notice. People generally understand that a visible camera observes a space. They may not expect an access point and receiver to classify objects beneath clothing.

That mismatch creates a consent problem. A public deployment would require clear rules covering notice, permitted uses, retention, access, and review. It would also need procedures for disputing an alert.

False positives carry more than an inconvenience in this context. A system that labels someone as a possible handgun carrier can prompt a security response. The consequences depend on how operators interpret confidence scores and combine them with other evidence.

The current research does not justify autonomous intervention. WiFireNet should not determine intent, legality, or identity. It only attempts to classify patterns associated with carried objects inside its experimental domain.

The distinction is particularly important in places where lawful concealed carry exists. A radio prediction cannot determine whether someone has authorization. It also cannot infer why an object is present.

Bias testing will need to extend beyond conventional demographic categories. Radio propagation can change with body size, clothing, accessories, mobility aids, carried electronics, and movement patterns. Uneven representation could produce uneven alert rates.

The four recorded object classes represent a thoughtful starting point, but public environments contain far more variation. Umbrellas, tools, cameras, battery packs, medical devices, laptops, and dense groups can create unfamiliar radio signatures.

Adversarial behavior presents another unanswered question. A person might change the handgun’s orientation, position, covering material, or proximity to other metal objects. The paper does not evaluate deliberate attempts to evade detection.

Infrastructure operators must also consider cybersecurity. A sensing system depends on trusted CSI streams, synchronized devices, trained weights, and alert software. Manipulated measurements or compromised models could suppress genuine detections or manufacture alerts.

Open-source release supports defensive evaluation. Researchers can test alternate splits, inspect preprocessing, reproduce model ablations, and search for failure cases. Independent teams can also determine whether the published scores survive different hardware and rooms.

The same openness lowers the barrier for experimentation outside formal institutions. That does not make release inherently irresponsible, but it increases the importance of governance alongside technical validation.

The appropriate near-term model is assisted sensing. WiFireNet might eventually provide another signal to trained personnel, alongside cameras, physical screening, and contextual judgment. It should not be treated as an invisible automated accusation system.

Three Tests Will Decide Whether WiFireNet Leaves the Lab

The next phase needs harder validation, not merely a larger score from a newer wireless standard.

The first signal to watch is a participant-held-out and session-held-out evaluation. Researchers should train on some people and recording sessions, then test on entirely separate ones. This would reduce the chance that neighboring samples or repeated trajectories inflate performance.

A strong result under that split would support the claim that WiFireNet learns object-related patterns. A steep decline would show that the model relies heavily on participants, movements, or recording conditions seen during training.

The second signal is room-held-out testing. A useful experiment would train in several rooms and evaluate inside another building without fine-tuning. Furniture, wall materials, transmitter placement, crowd density, and unrelated wireless traffic should all vary.

This test targets the system’s primary opponent: environmental dependence. Stable performance would strengthen the case for transferable Wi-Fi sensing. Failure would limit the technology to calibrated, site-specific installations.

The third signal is an operational error profile from larger crowds. Researchers need to report false alerts per person or per hour at selected thresholds. They should also publish miss rates for different weapon positions, clothing types, and movement patterns.

That evidence matters more than a single average-precision headline. Security operators make threshold decisions, and those decisions create practical consequences. A model with reasonable benchmark performance can still become unusable if alerts arrive too frequently.

Future experiments can also examine Wi-Fi 7 or emerging Wi-Fi 8 capabilities. Wider channels and additional spatial measurements might improve resolution. The current study offers no basis for assuming that outcome, because WiFi-plus already produced the opposite result.

Independent reproduction should come before hardware optimism. The released model and data allow outside teams to reconstruct the published pipeline. The more valuable step is collecting new datasets with different devices, rooms, and participants.

WiFireNet handgun detection has crossed an important research threshold. It detected evidence associated with concealed handguns in dynamic, multiperson indoor scenes using only CSI at inference.

It has not crossed the deployment threshold. The system was tested in two small rooms, against four object classes, with sample-level data splits. Its strongest number also weakened under stricter localization.

Developers and security researchers should now ask whether the signal survives changes the original model never saw. Can it generalize across buildings, hardware, people, clothing, clutter, and larger crowds without constant recalibration?

Those tests will determine whether WiFireNet becomes a useful secondary sensor or remains an instructive laboratory benchmark. Until then, its 72.5% result should open a research program, not a procurement process.