KAIST's Devilray rebuilt 2,592 fake cell tower setups, and 7 detectors each missed 192+

Researchers at KAIST have built Devilray, a test platform that reproduces 2,592 distinct behaviors of fake base stations, the rogue cell towers also known as IMSI catchers or StingRays. When the team ran it against seven publicly available detectors, every detector missed at least 192 of those configurations, according to the university.

The point is less any single gap than how the field has been checking its own work. For about two decades, the KAIST group argues, detectors have mostly been tested against fake towers that researchers built themselves, so a high score may only mean a detector recognizes its own design.

A commercial unit as the starting point

The team began with a fake base station sold to law enforcement and intelligence agencies, which KAIST describes as the first such device analyzed in an academic study. Such units are expensive and their sale is restricted, which is why university labs usually cannot get one.

From that device, the researchers worked outward through earlier research and the 3GPP cellular standards. They broke an operation into four steps (scanning the nearby network, standing up a fake cell, drawing phones to it, then the attack itself), identified seven dimensions an attacker can vary, and applied 21 rules to drop combinations that could not occur together. Every one of the remaining 2,592 configurations was run against real smartphones to confirm it completes.

The researchers stress that the figure does not describe attacks observed in real life. It is a reference adversary, built from a real device and from what the standard allows.

Why LTE leaves room to maneuver

The weaknesses are not bugs that a patch can remove. An LTE phone does not check whether a base station is genuine when it first connects, it acts on some messages before authentication, and it can be forced back onto earlier network generations. Because this is written into the standard, an attacker can stay fully compliant while choosing among many routes to the same goal.

One example from the study: a fake tower can ask a phone directly for its permanent identifier, which is the request the Electronic Frontier Foundation's open-source Rayhunter detector looks for. Or it can reject the phone, prompting it to reconnect and volunteer the identifier unprompted. Other detectors flag the strong signal of a cheap fake tower and miss one that lowers its output until it reaches only a few meters.

The catch in simply retraining detectors

The team then fed a machine learning detector the Devilray cases that had slipped past it and retrained it. On its radio link layer model, detection rose from 51% to 96%, but false alarms on ordinary carrier traffic went from 1.4% to 5.3%. On the core network model, detection climbed from 27% to 63% while false alarms rose from 4.7% to 8.8%, reaching 15% when trained on the hardest case.

"The real question is whether we can reduce misses without raising false alarms," said Yongdae Kim, the KAIST professor who led the work. The authors argue that future detectors should always report both numbers together.

Access to Devilray is restricted to screened researchers and organizations. According to KAIST, the team has given EFF the cases Rayhunter misses, and the paper will be presented at the ACSAC 2026 conference in Los Angeles, held December 7 to 11, 2026.