A hidden camera detector you can build for the price of two coffees, that catches a spy camera whether or not it's even switched on — that's the pitch behind SweepLED, and it's why the story shot to #1 on Hacker News (141 points, 38 comments) within a day of Chosun Ilbo's August 30-31, 2026 coverage of the research.
SweepLED comes from Professor Han Jun's team at KAIST's School of Computing, working with researchers at NUS (National University of Singapore) and the Singapore Management University. It was presented at ACM MobiSys 2026 in June — this is peer-reviewed academic research, not a startup pitch — and the core claim is straightforward: a $7 LED clipped to a smartphone, paired with an AI model, spots hidden cameras in under 5 seconds at 94% accuracy.
TL;DR
| Question | Answer |
|---|---|
| What is it? | A smartphone-attachable LED device plus AI model that detects hidden cameras by analyzing reflections |
| Who built it? | KAIST (Prof. Han Jun's team), with NUS and Singapore Management University |
| Where was it presented? | ACM MobiSys 2026, June 2026 |
| How accurate? | 94% across 30 tested household objects |
| How fast? | Under 5 seconds per scan |
| Hardware cost | ~$7 (10,000 Korean won) for the core LED component |
| Does it work on powered-off cameras? | Yes — it reads lens optics, not activity |
| Can I buy one? | Not yet — it's a research prototype, not a shipping product |
| Does it beat every countermeasure? | No — lensless/computational sensors are a theoretical blind spot, and it's a close-range scan |
Why existing hidden camera detectors fail
Portable hidden-camera detectors already exist, and most work the same way: shine a light (often IR) at a suspicious object and look for a bright reflected dot where a lens would be. The problem is that plenty of ordinary things — metal trim, glass, glossy plastic — throw back a bright reflection too. Users end up second-guessing themselves over a shiny screw or a phone charger's glass window, which is exactly the accuracy problem SweepLED's paper cites as its starting point.
The single-bright-spot method has a fundamental flaw: it only checks whether something reflects light, not how it reflects light differently from real optics.
The core insight: reflections change shape differently for real lenses
SweepLED's approach keeps the smartphone's own camera fixed on the target object, while the LED continuously sweeps its direction of illumination. As the light angle changes, the device captures how reflections on the object's surface move, shrink, or disappear.
A flat glossy surface behaves simply under a sweeping light: the reflection point slides across the surface or vanishes once the angle is wrong, the way a highlight moves across a polished tabletop. A real camera lens does something structurally different. Because a lens is built from multiple internal lens elements, an aperture, and an image sensor stacked behind it, illuminating it from different angles produces a distinctive, layered reflection pattern that shifts in position and shape in a specific way tied to that internal structure — not a single glint, but a signature.
A deep learning model trained on this behavior analyzes the entire pattern of changes across multiple light directions, rather than checking for one bright spot. In the team's tests across 30 common household objects — chargers, alarm clocks, remote controls, decorations — that full-pattern approach hit 94% accuracy, each scan finishing in under 5 seconds.
As Professor Han Jun put it: "Illegal hidden cameras seriously threaten personal safety and privacy in everyday spaces... By combining low-cost hardware with AI analysis, we have demonstrated the potential for a detection technology that even non-experts can use easily."
This is the same underlying shift that's reshaped a lot of practical computer vision lately — moving from single-cue heuristics toward models that read a full pattern of signal across conditions. It's the same reasoning behind why modern detection architectures like RF-DETR outperform older single-pass detectors: more context, evaluated holistically, beats one cheap heuristic.
Why this beats motion, heat, and RF detection: it works on cameras that are off
The Hacker News thread zeroed in on the detail that matters most for anyone actually worried about being watched: does the method still work if the camera is switched off?
User carlostkd raised the obvious countermeasure — a camera that stays dark for hours after a room is scanned, only activating once the coast looks clear, would defeat any detector that relies on catching the camera in the act. That defeats motion detectors, heat signatures, and RF-signal sniffers, all of which need the camera doing something.
User zack6849 correctly pointed out why SweepLED doesn't have that weakness: its method analyzes the lens's physical characteristics through reflection, which "works regardless of whether the camera is powered on or off, since it's detecting the optical structure of the lens itself, not its activity." A lens is a lens whether or not current is flowing to the sensor behind it — the glass elements and aperture are still there, still reflecting light the same distinctive way. That's a structural advantage reflection-based detection has that activity-based detection can never fully close.
The infrared thread: why your phone can't just "see" hidden cameras
A long, genuinely technical sub-thread on Hacker News (contributors including baking, hammock, superxpro12, TheRealPomax, and ChrisMarshallNY) worked through a common assumption: can't you just point a phone camera at something and look for an IR glow?
The consensus that emerged: most modern phone camera lenses have IR-blocking filters built in — visible, if you know to look, as a blue-tinted glass rather than the black IR-pass glass used on cameras that have been deliberately converted for infrared work. Some older or cheaper phones, and some front-facing cameras, skip a strong IR filter, so results vary by device. But the more important asymmetry, per the thread, runs the other way: surveillance cameras themselves often deliberately lack IR filters, because they need infrared sensitivity for night vision.
Commenter evilduck sharpened that point: a hidden camera that includes its own IR illuminator (for night recording) is actually easier to spot than a SweepLED-style scan, because an IR-sensitive camera — even an imperfect stock phone one — can show a visible glow right at the illuminator, no reflection-angle analysis required. IR-emitting spy cameras give themselves away for free; the harder case is a camera with no IR illuminator and no obvious tell, which is exactly the case SweepLED targets.
The laser-scanning alternative, and its own honest gaps
User kragen brought up the established alternative technique: scanning a low-power (10mW) 832nm laser across a room with a mirror, a method popularized in a well-known Dan Gelbart lecture on optics and security countermeasures. The same retroreflection principle that lets you spot a camera lens this way also works for detecting eyeballs and night-vision equipment in a dark room, since both reflect a narrow beam straight back at the source in a telltale way.
kragen also raised the deeper theoretical concern that applies to any reflection-based detector, SweepLED included: lensless imaging — compressed sensing or Hadamard-mask-based sensors that reconstruct an image computationally rather than through a stack of glass lens elements — wouldn't produce the reflection signature these methods rely on, in principle defeating detection entirely. That's a research-lab class of sensor today, not something in common spy cameras, but it's the honest theoretical ceiling on this whole family of technique. kragen also noted the simplest countermeasure of all against any close-range scan: watching from far enough away that nobody's running a phone-camera sweep on you in the first place.
SweepLED shares that last limitation squarely — it's a close-range scan, the same as holding a phone up to an object, not a room-scale sweep. It won't catch a camera positioned across a space the way a laser-mirror scan can.
Why this matters beyond "creepy tech story"
A recurring HN sub-thread (contributors including dylan604, expedition32, and dbspin) pushed back on framing this as niche paranoia. The real audience isn't just "an adult male worried about being filmed" — it's families renting short-term stays with kids, people checking AirBnBs and vacation rentals, and anyone in a space they don't fully control. A $7 sensor and a five-second scan is cheap enough to make "check the room" a routine habit rather than a specialist tool, which is the practical case for why low-cost, high-accuracy detection matters more than the lab number alone suggests.
What to do after you've cleared the room
Finding (or ruling out) a hidden camera solves one half of a privacy problem. The other half shows up the moment you actually share photos or video from a space that had other people in it — a rental, an event, a public area. Detecting cameras doesn't retroactively give you consent to publish whoever else is in frame.
This is where a tool like bgblur.com fits as the natural next step, the same way we've covered it as a consent-respecting sharing layer for self-hosted photo libraries: browser-based AI face detection that blurs bystanders before you post, with no software install and no watermark. If you're sharing vacation photos from a rental you just swept for cameras, or clips from an event where SweepLED-style checks turned up nothing, blurring faces you don't have consent to publish is the other side of the same privacy discipline — see our dedicated guides on blurring faces in photos and blurring faces in video for the step-by-step workflow.
It's the same broader privacy-tech space we've been tracking on this blog with the rise of always-on capture devices — see our coverage of smart glasses misuse and venue bans for the flip side of the same question: not "is someone watching me," but "am I the one recording someone else without asking."
Honest limitations, in one place
SweepLED is a strong, well-documented research result — not a magic guarantee. Worth being precise about what it does and doesn't claim:
- 94% accuracy is a lab number on 30 tested household objects, not a certified rate against every hidden camera model sold today.
- It's a close-range scan. Like any phone-camera check, you need to hold the device near the object you're inspecting — it won't sweep a whole room from a doorway.
- Lensless/computational sensors are a theoretical gap. Compressed-sensing cameras with no conventional lens stack wouldn't produce the reflection signature the model looks for — rare in practice today, but a real limit of the technique.
- It's not commercially available yet. This is peer-reviewed academic work presented at ACM MobiSys 2026, with no announced consumer product or ship date.
None of that undercuts the core contribution: proving a $7 LED component plus a trained model can beat the accuracy of existing detectors that cost more and require the user to spot a single bright dot by eye. That's a meaningfully lower bar to clear for anyone who wants to check a room without specialized gear.
Related on explainx.ai
- Self-host your photos: Immich 3.0, privacy, and bgblur.com
- How to blur a face in a photo (free AI tool)
- How to blur faces in videos with AI
- RF-DETR: Roboflow's real-time detection transformer, explained
- Smart glasses misuse, venue bans, and how to protest
- EU driver-facing camera law: privacy guide
- Digital camouflage vs AI surveillance
External references
Research details, accuracy figures, and hardware cost reflect the ACM MobiSys 2026 paper and Chosun Ilbo's August 30-31, 2026 coverage. SweepLED is a research prototype as of this writing, with no announced commercial release — verify current availability before assuming a consumer product exists.