Investigation of GPS interference events with

refinement on the localization algorithm

Zixi Liu, Stanford University
Juan Blanch, Stanford University
Sherman Lo, Stanford University
Todd Walter, Stanford University

BIOGRAPHY
Zixi Liu is a PhD candidate at the GPS Laboratory at Stanford University. She received her B.Sc. degree from Purdue University
in 2018 and her M. Sc. degree from Stanford University in 2020.
Juan Blanch is a senior research engineer at the GPS laboratory at Stanford University.
Sherman Lo is a senior research engineer at the GPS laboratory at Stanford University.
Todd Walter is a Professor of Research and director of the GPS laboratory at Stanford University.

ABSTRACT
GNSS serves safety-of-life applications in aviation such as precise navigation for approach and landing operations. Interference
events happen near airport can severely affect the safe operations of the airspace. A recent interference event happened at
Dallas-Fort Worth International Airport (KDFW) in October/2022 caused a widespread disruption. This incident resulted in
multiple aircraft reporting GPS unreliable within 40NM, closure of a runway, and rerouting of air traffic. In this study, we
performed a detailed investigation on this event, and run our localization algorithm to provide an initial estimation of the
potential jamming source.
There were no public reports from ground infrastructures during this event, which means collecting data from the ground is
not sufficient. Therefore, in this study, we used data collected from Automatic Dependent Surveillance—Broadcast (ADS-B)
system. It is a satellite-based surveillance system on the airplane which broadcasts aircraft position information. ADS-B is
already widely in use and was made mandatory in Europe and the U.S.A. by 2020. This ubiquity and openness of ADS-B
provides widely available source of GNSS information.
In addition to investigating Dallas event, this research also built on our previous work on localizing interference sources (Liu
et al., 2022) and provided a method to calculate an error bound on the final estimated jammer location. In our prior research, we
built an algorithm that can identify the most likely location and transmitted power of potential jammer in real time. In this work,
we designed an algorithm to provide real-time confidence information about the localization result. The error bound calculated
from this confidence monitoring scheme is compared with result from the bootstrap method (Stine, 1989). The goal of this
design is to help narrow down the ground searching area in order to physically shut down the jamming source. We implemented
and demonstrated this capability using recorded ADS-B transmissions from known interference events.

I. INTRODUCTION
Automatic Dependent Surveillance—Broadcast (ADS-B) is a surveillance system on the aircraft that periodically broadcasts
the GNSS derived position every 0.4 – 0.6 sec on a 1090MHz frequency band (DO-260B RTCA (Firm), 2011). The benefit
of using ADS-B data for interference detection and localization, compared with traditional methods such as radio direction
finding, is that it’s less time consuming. In addition, this crowd-sourced data from multiple aircraft offers broader coverage
of the impacted airspace. The drawback to this method is that ADS-B message does not contain information from the GNSS
receiver such as carrier to noise ratio (C/No). However, ADS-B messages have some built-in parameters which indicate the
integrity and accuracy levels of the reported GNSS measurements. In our research, we used Navigation Integrity Category
(NIC) to estimate the quality of GNSS reception. NIC describes the size of an integrity containment radius that the current
horizontal position is guaranteed to be within, with 99.999% probability. Figure1 shows the tables for NIC and how it reacts
under different circumstances. According to ADS-B equipment performance requirements defined in Title 14 of the Code of
Federal Regulations (CFR) Part 91 by Office of the Federal Register, National Archives and Records Administration (2013),
under normal circumstances, the aircraft’s NIC value must be ≥ 7 (containment radius less than 0.2 nautical miles). When NIC
equals to 0, the corresponding errors are much worse than typical GNSS performance. For GNSS to perform this poorly, other
than incorrect setting of the ADS-B system, it is likely that the aircraft has been severely affected by the jammer.

Figure 1: Relation between NIC and size of containment radius.

In this study, we used ADS-B data to investigate an interference event happened at Dallas-Fort Worth International Airport
(KDFW) in October/2022. We refined and applied our localization algorithm which identified the most likely location of the
jammer with an error bound for the Dallas event. In addition, it generated a probability heat-map for the possible locations
of the interference source. The main goal of this research is to design a system of algorithms that is able to perform rapid
detection and accurate localization of GNSS interference source. The resulted information should help identify and characterize
some properties about the potential jamming source. The ultimate goal is to have this algorithm be able to quickly identify and
guarantee finding the interference threat in order to physically shut down the jamming source.

II. RELATED WORK

Several groups have investigated the use of ADS-B on detection and localization of GNSS interference events. Aireon is able to
provide alerts of potential GPS interference events by monitoring change of Navigation Accuracy Category–Position (NACp)
parameter from ADS-B message (Garcia, 2020). EUROCONTROL has investigated the use of ADS-B to determine GNSS
affected regions in the eastern Mediterranean. They developed a grid probability model, based on ADS-B trajectory gaps, to
identify possible location of the radio frequency interference (RFI) source (Jonáš and Vitan, 2019).
The work in this paper builds on our prior real-time interference detection and localization work (Liu et al., 2022) and expands
that capability with more information on the reliability and robustness of that algorithm. The main concept of our previous work
relies on calculating the minimum least squares solution of jammer characteristics such as location and transmitted jamming
power. In this paper, we analyze the underlying properties of ADS-B reports such as the nominal noise of the measurements,
the uncertainty in the localization result, and we improve our localization algorithm by incorporating the variance of the noise
into the objective function.

III. DATASET
In this paper, we validated and tested the algorithm on two GPS interference events. The first event occurred around Denver
International Airport (KDEN) in January 2022. Figure2 shows how that interference event affected aircraft GNSS reception
and hence ADS-B outputs. On the left is a top view picture of aircraft flight tracks passing KDEN during few hours on regular
days, and the plot on the right shows data collected from few hours during the interference. The color of each dot represents
corresponding NIC value. On regular days, all aircraft have NIC value greater than or equal to 7. On the contrast, during
interference, some aircraft have low NIC values such as NIC = 0 (dark blue) which can be observed at the center of the plot on
the right.
 Figure 2: Top view of ADS-B data collected from KDEN before (left) and during(right) interference event.

The second event happened recently around Dallas-Fort Worth International Airport (KDFW) in October/2022. This event
caused dramatic impact including a closure of a runway, and rerouting of Texas air traffic. The cause or the source of the
interference is still unknown at current stage, we did our own investigate on this event and we run our algorithm to estimate the
most likely jammer location with a probability heat-map. Figure3 shows a sample plot of how aircraft has been affected during
a 1-hour time window. To have a better visualization, all position points with low NIC values (NIC < 6) are marked in red and
position points with high NIC values (NIC ≥ 6) are marked in green. The detailed investigation of this event is shown in the
result section.

Figure 3: Top view of ADS-B data collected from KDFW airspace on October/17/2022.

IV. METHOD
The localization algorithm we designed involves solving linearized weighted least squares problem of the jammer characteristics
which are location and transmitted jamming power. Therefore the unknown states of the algorithm can be written as [x, y, z, Pt ]
in Earth-centered, Earth-fixed coordinate system (ECEF).
The measurement in this problem is the amount of jamming power received by the aircraft at each position point which is
written as Pr [dBW ]. Since ADS-B data does not provide direct information about received jamming signals, we need to infer
information about the GNSS reception from the NIC value of ADS-B data. We characterized the relation between NIC and
received jamming power through data collected from a Trimble receiver under attack of wideband jammer. Figure 4 shows
the experiment result. The x-axis is the received jamming power at the aircraft antenna and the y-axis is the NIC value. The
experiment is done by simulating transmission of different levels of jamming power to the aircraft antenna. For each power level,
we observed a set of position points with protection level information. Then we convert protection level to NIC value based
on table shown in Figure 1. The result gives us an observable relation between received jamming power and NIC value. The
top plot shows box plot of the experiment result. The bottom plot shows how many percentage of points were at corresponding
NIC value and the mode value at each power level is connected by a black line. This experiment helps us understand what is the
expected amount of received jamming power for a given NIC value and the corresponding variances.

Figure 4: Relation between received jamming power and NIC value.

To estimate or predict the amount of jamming power received at each point, we use the Friis formula (H. T. Friis, 1946) which
shows the power loss of a signal from the transmitter to the receiver in a space without signal blockage. Pr is the power
received in watts, Pt is the power transmitted in watts, λ is the transmitted signal wavelength in meters, d is the distance between
transmitter (jammer) and receiver (aircraft) in meters, Gt , Gr are the transmit and receive antenna gain.

Gt Gr λ2 Pt
P̂r =
(4π)2 d2 (1)
where d2 = (x − xa )2 + (y − ya )2 + (z − za )2

The objective function then becomes to minimize the difference between the estimated (Pˆr ) and measured (Pr ) received jamming
power at each position point. The aircraft position can be written as [xa , ya , za ] in ECEF:

minimize (Pr − Pˆr )T W (Pr − Pˆr ) (2)

W in the above equation is a weight matrix that allows us to design the mechanisms behind the model. The weight matrix in
our algorithm is the inverse of the covariance matrix. The diagonal elements of the covariance matrix represent the variance
from the measurement noise. The measurement noise can come from different sources such as during the process of inferring
received jamming power from NIC value. The variance of the measurement noise can be estimated from the residuals:

1
σ̂ 2 = ∥Pr − Pˆr ∥2 (3)
m−n

Where m is the number of measurements and n is the number of unknown parameters in jammer state. Once we have the variance
of the noise, we can calculate the error variance of the estimator from the covariance matrix of the least squares estimator. One
thing special about this problem is that the objective function itself is non-linear, to better fit the problem into weighted least
squares problem, we first convert power from watts to dBWatts by taking 10log10 of both sides, then linearize the objective
function about current Kth iteration:

Df
f (X) = Pr − P̂r ≈ f (X (k) ) + (X (k) )(X − X (k) ) = A(k) X − b(k)
h DX i
where A(k) = 2(x−x a) 2(y−ya ) 2(z−za ) −1 (4)
ln(10)d2 ln(10)d 2 ln(10)d 2 ln(10)Pt

where b(k) = A(k) X (k) − (log10 Pr − log10 P̂r )

Then the 95% confidence interval of ith estimator in jammer state X = [x, y, z, Pt ] can be calculated by:

p
x̂i ± 1.96 var(
ˆ x̂i )
(5)
ˆ x̂i ) is ith element on the diagonal of (A(k)T A(k) )−1 σ̂ 2
where var(

V. RESULTS AND ANALYSIS

1. Investigation of Dallas GPS Interference Event
A recent GPS interference event happened at Dallas on 10/17/2022 caused dramatic impact and widespread disruption. We
obtained ADS-B data from 10/17/2022 00:00:00UTC to 10/20/2022 23:59:59UTC. During the investigation, we noticed few
types of anomalies that contain misleading information about the interference in terms of impact region and timeline. First type
of anomaly is military trainer aircraft which were doing high dynamic maneuvers and remaining in the patterns as shown in
Figure 5. These unusual rapid flight maneuverings can lead to a loss of GPS signals on the aircraft which causes the onboard
ADS-B system to report low NIC values. These low NIC values can contain misleading information as those regions seem to
be affected but in fact no interference exists. Other aircraft which were flying nearby have only been reporting good NIC values
and showed no evidence on the existence of interference. Second type of anomaly is aircraft reporting low NIC values along
the entire flight path regardless of being affected by the interference as shown in Figure 6. This could happen if the aircraft has
low versions of ADS-B system or incorrect settings. Normally we observe decrease in NIC value only when the aircraft is close
to the impact region and the NIC value should recover back to value ≥ 7 after the aircraft has left the impact region. The third
type is flight with NIC value jumping back and forth, such as the flight shown in Figure 6, it has NIC value changes between 0
and 9 during entire flight path. This type of sudden decrement in integrity level followed with an immediately recovery within
one second is unusual. The final type is non-realistic side jump in flight path as shown in Figure 6. The change of position from
one location to another location is non-realistic when those two positions are too far apart and the time different is only within
few seconds. This type of behavior is a clear indication of incorrect position information.
 Figure 5: ADS-B position points with Low NIC values due to high dynamic maneuvers of military trainer aircraft.

Figure 6: Top view of three commonly seen types of anomalies.

By removing the above types of anomalies, we now have a better understanding of the timeline and impact regions of the Dallas
event. Figure7 shows how many numbers of flights have been affected with respect to time. We identified jammed flights based
on the reported NIC value and the flight tracks behaviours such as sudden loss of the flight path. The significant impact of
the jamming started at around 10/17/2022 19:21:00UTC and ended at around 10/18/2022 19:10:00UTC. Figure8 shows the
overall picture of how aircraft have been affected spatially. For better visualization, this plot only shows flight paths observed
during a 3-hr time window. On the left is a top view of all flight trajectories, on the right is the side view of flight paths with
respect to latitude and longitude. By looking at only the points that were affected, the impact region seems to indicate that the
jammer is more likely to be a directional antenna than omnidirectional, and this antenna type information is considered during
the implementation of our localization algorithm.
 Figure 7: Numbers of flights been affected before, during, and after the jamming event.

Figure 8: Top view (left) and side view (right) of all flight tracks during 3-hr time window.

2. Error Variance of Weighted Least Squares

In this study, we provided additional error bound to our localization algorithm. To validate the design, we used data collected
from a known interference event, and we compared with results from the bootstrap method. Figure9 shows the histogram of
residuals and the corresponded fitted distribution using data collected from interference event happened at KDEN. The variance
of measurement noise is σ 2 = 53.1528[dBW 2 ]. Using equation 3, 4 and 5, the calculated error variance of the jammer location
and power which is also the diagonal elements of the covariance matrix is [1.5819e + 07, 1.1975e + 08, 9.0713e + 07, 3.382].
Recall that these variance correspond to the location of the jammer (in ECEF) in meters and transmitted jamming power in
Watts.
 Figure 9: Histogram and fitted distribution of the measurement noise.

Figure10 shows the results obtained from Bootstrap method. We created 10,000 sets of data. The plot on the left shows jammer
locations calculated from using bootstrap method and the cyan cirle shows the jammer location calculated from running Gauss
Newton method on the original dataset. The plot on the right shows distribution of estimated jammer latitude and longitude.
Recall that in the least squares method, we represent jammer location in terms of (x, y, z) in ECEF. Therefore, we need to
convert latitude and longitude from degree to meters in ECEF. The resulted error variance from the 10,000 subsets of solutions
is [1.3798e + 07, 8.1378e + 05, 2.7101e + 06, 1.360].

Figure 10: Estimated jammer locations and distribution from Bootstrap method.

Now that we have the error variance from both methods, we can then calculate the corresponding error bounds with 95%
confidence interval using equation 5. Figure 11 shows the error bound from both methods. The error bound obtained from
least squares is larger than the error bound from bootstrap method. This is as desired since the algorithm needs to protect the
reliability by having more robust bound on the estimated jammer location.
 Figure 11: Error bound from Least squares and Bootstrap method.

3. Localization results for Dallas event

Now that we have the localization algorithm designed and validated, it is time to run the algorithm on the interference event
happened in Dallas. We run our algorithm on a snapshot of ADS-B data collected from 3-hr time period during the GPS
interference event.
a) omnidirectional antenna
We first assumed that the interference source is static omnidirectional with continuous transmission and no signal blockage
by terrains. Figure 12 shows the heatmap of residuals at each possible jammer location. Red indicates smaller residuals and
therefore lower region on the surface plot of the objective function. Since we have 4 variables in this problem, to visualize the
objective function in 3D, at each possible jammer location, we only plot the residual corresponds to the Pt value that gives
minimum residual in compared with other Pt values. In other words, at each possible jammer location, only the most likely
transmitted jamming power is considered. This value of Pt is written as Ptopt . The visualization of the objective function gives
us better understanding of the overall jamming situation in the selected airspace.
The cyan star shows the global minimum point on the residual plot. From 20:00UTC to 23:00UTC, the most likely jammer
location is at (-96.39, 32.18). The corresponding most likely transmitted jamming power Ptopt is about 100 watts. One thing
we noticed is that the most likely jammer location varies as time changes, therefore, to better estimate this jammer location, we
might need to incorporate non-static or non-continuous transmission jammer model into our localization algorithm.
The red star and the red circle together shows the final estimated jammer location with an error bound. This is obtained by
running Gauss Newton method on an initial assumed jammer location for multiple iterations until converged. Benefit about
this algorithm is that during real-time implementation, it is less time consuming than generating a overall heatmap on the entire
selected airspace. This selected airspace has an area of 300NM x 460NM and the locations of some major cities are shown on
the map. The drawback is that the final estimated result is close to but not right on the global minimum point.
This heatmap seems to indicating two most likely regions that could contain the potential jamming source, and based on the
investigation on the ADS-B data collected from the airspace as shown in Figure8, the jammer is more likely to be directional.
Therefore, our next step is to change the assumed model of the potential jamming source.
Figure 12: Residual heatmap with the localization result and corresponding error bound, assuming transmitter with omnidirectional antenna.

b) directional antenna
We can change the antenna type from omnidirectional to directional by changing the antenna gain pattern which is Gt in
equation2. By looking at only the affected points as shown in Figure13, it is likely that the jammer is a directional antenna
pointing towards an azimuth angle aiming 30◦ or 210◦ . This direction does not align with any runway at Dallas-Fort Worth
International airport, so it is also possible that the interference source is omnidirectional but the jamming signal is blocked from
the east and west side by nearby buildings or mountains.

Figure 13: Sampled ADS-B data with an emphasis on affected position points (low NIC).

Figure14 shows the residual heatmap of jammer location assuming it is static directional with continuous transmission. This
objective function points more strongly to the global minimum point which means the assumption of directional antenna fits
better to current dataset.
 Figure 14: Residual heatmap with the localization result and corresponding error bound, assuming transmitter with directional antenna.

VI. CONCLUSION AND FUTURE WORK

In this study, we identified the potential misleading information that could come with low NIC values. We associated non-
standard low NIC values with type of aircraft such as military trainer aircraft. By filtering out those anomalies, we then improved
the understanding of Dallas event in terms of timeline and impact region. We identified that the antenna type of the potential
jamming source is more likely to be directional rather than omnidirectional.
We refined our localization algorithm by providing additional uncertainty and confidence information on the localization result.
The error bound estimation from variance of measurement noise provides immediate and better error bound compared with the
error variance obtained from bootstrap method. The implementation of the algorithm onto the recent interference event shows
a potential of identifying the most likely location of the interference source and the error bound provides an extra uncertainty
information on the localization result. In the next step, we would like to further investigate the Dallas event and collect more
real-life data to test and improve our algorithms. In addition, we will incorporate different types of antennas as well as the
non-static transmitter into the localization algorithm. As those types of sources are more likely to be the interference source of
the Dallas event. Furthermore, we will also consider effects of flight configurations and aircraft fuselage on jamming signals in
order to have more realistic modeling on the measurement noise.

ACKNOWLEDGEMENTS
We gratefully acknowledge the support of the FAA Satellite Navigation Team for funding this work under Memorandum of
Agreement #: 693KA8-22-N-00015. We also thank OpenSky Network(Schäfer et al., 2014) and ADS-B Exchange for providing
ADS-B data for this study.

REFERENCES
DO-260B RTCA (Firm) (2011). Minimum Operational Performance Standards (MOPS) for 1090 MHz Extended Squitter Auto-
matic Dependent Surveillance-Broadcast (ADS-B) and Traffic Information Services-Broadcast (TIS-B). RTCA, Washington,
DC, USA.
Garcia, M. (2020). Global surveillance and tracking of aircraft via satellite. presented to SCPNT Symposium, Stanford
University, CA, USA. [Powerpoint Slides].
H. T. Friis (1946). A note on a simple transmission formula. Proceedings of the IRE, 34(5):254–256.
Jonáš, P. and Vitan, V. (2019). Detection and localization of gnss radio interference using ads-b data. In 2019 International
Conference on Military Technologies (ICMT), pages 1–5.
Liu, Z., Lo, S., Walter, T., and Blanch, J. (2022). Real-time detection and localization of gnss interference source. In Proceedings
 of the 35th International Technical Meeting of the Satellite Division of The Institute of Navigation (ION GNSS+ 2022), pages
3731–3742.
Office of the Federal Register, National Archives and Records Administration (2013). 14 CFR § 91.227 - Automatic Dependent
Surveillance-Broadcast (ADS-B) Out equipment performance requirements. [Government].
Schäfer, M., Strohmeier, M., Lenders, V., Martinovic, I., and Wilhelm, M. (April 2014). Bringing up opensky: A large-scale
ads-b sensor network for research. In ACM/IEEE International Conference on Information Processing in Sensor Networks.
Stine, R. (1989). An introduction to bootstrap methods: Examples and ideas. Sociological Methods & Research, 18(2-3):243–
291.