Reading — Modern Trends in EW#
By the end of this lesson you should be able to:
Describe the cognitive-EW loop — sense, learn, adapt — and explain why a fixed threat library fails against agile, reprogrammable emitters.
Read a machine-learning emitter classifier’s confusion matrix as a probability of detection \(P_d\) and a probability of false alarm \(P_{fa}\), and place it on L8’s ROC.
Compute a GPS denial radius from a jammer’s power, free-space path loss and the C/A code’s processing gain, and the INS drift accumulated over a GPS-denied leg.
Outline space-EW geometry: why uplink jamming needs the satellite’s receive beam and downlink jamming is local.
The playbook, and the assumptions under it#
L31 through L36 built a survivable route out of known pieces. You put a probability on every link of the kill chain, drew each site’s detection contour and missile engagement zone, weighed how well a passive network could track you, searched a risk graph for the best route, localized a pop-up threat, and bought a standoff jammer and a SEAD strike to open a corridor. It is a complete method — and every piece of it rests on assumptions you have not had to question: that the threat parameters in your laydown are the real ones, that an identification is a fact, that you always know where you are, and that the links coordinating your support package will be there when you need them.
This lesson surveys four trends that are already attacking those assumptions. Each is a field of its own; here each gets the one idea you need to reason about it, and one of them — navigation warfare — gets a number you can compute.
Trend |
The assumption it breaks |
|---|---|
Cognitive EW |
the threat library is complete, and the threat behaves as catalogued |
AI/ML emitter identification |
a classifier’s label is a fact, not a probability |
Navigation warfare (NAVWAR) |
you always know where you are — GPS is truth |
Space EW |
the satellites and links behind your support package are always there |
The right-hand column is also where the Limitations section of your Final Project comes from. The paper asks you to pick one of these trends and say what it would change in your route; the last section of this reading shows what a good answer looks like.
Cognitive EW: sense, learn, adapt#
Go back to L12. A radar-warning receiver deinterleaves the pulse stream into emitters, measures each one’s parameters — frequency, pulse repetition interval, pulse width, scan — and matches them against a threat library. The library is the product of electronic intelligence (ELINT): someone collected the emitter, analyzed it, and wrote down what it looks like in each of its modes. L12’s type-along already showed the weak joint: a library that held one PRI per emitter stopped matching the moment the tracking radar changed mode, which is why real libraries carry parameters per mode.
Now take that one step further. A modern radar’s waveform is largely software. L15 showed you frequency agility and waveform diversity as electronic protection; push them to their end and you get an emitter that can present a mode nobody has ever catalogued — reprogrammed between engagements, or chosen on the fly. Against that emitter a lookup fails in one of two ways: it finds no match, or it finds the wrong one. And the fix — collect the new mode, analyze it, rebuild the library, load it into every aircraft — is a cycle measured in weeks, not seconds.
Cognitive EW is the attempt to close that cycle inside the mission. Haigh and Andrusenko’s survey (Cognitive Electronic Warfare: An Artificial Intelligence Approach, Artech House, 2021) frames it as a loop:
Sense — electronic support as before: intercept and characterize the emitter, its parameters and its behavior over time.
Learn — build or update a model of what the emitter does, without needing a library entry: which parameters it changes, how it reacts, what function it appears to serve.
Adapt — choose a response (a jamming technique, a maneuver, an emission-control change), then score it by watching what the emitter does next, and feed that score back into the loop.
The score is the new ingredient. A library lookup is open-loop: it names the emitter and cues a canned response. The cognitive loop measures whether its response worked — did the radar’s track mode break, did it change frequency, did it switch to a backup mode — and uses that to choose the next one.
Two cautions keep this honest. First, the loop runs on both sides. A cognitive radar senses your jamming and adapts its own waveform in reply; L15’s electronic-protection menu stops being a design-time choice and becomes an in-flight one. Second, a system that learns during the mission can learn the wrong thing, and it is hard to test and certify something whose behavior is not fixed in advance. Trust is the weak spot of the loop, as agility is the weak spot of the library.
Key Concept
A threat library matches what someone has already catalogued; a cognitive loop senses, learns and adapts in the mission, and scores its own responses by their effect. Agile, software-defined emitters defeat the library by presenting modes it has never seen faster than it can be reprogrammed — and the same loop, run by the adversary, lets their radars adapt to your jamming.
AI/ML emitter identification: read the confusion matrix#
Machine learning enters EW most directly at L12’s hardest step, classification. Instead of a hand-built lookup from measured parameters to an emitter name, a classifier learns that mapping from examples. However it is built, it is judged the same way: test it on intercepts whose true identity you know, and tabulate what it said. That table is a confusion matrix.
Here is a notional classifier tested on 100 intercepts of each of L7’s three radar functions — early-warning (EW), acquisition (ACQ) and target-tracking (TTR):
True emitter |
said EW |
said ACQ |
said TTR |
|---|---|---|---|
EW |
92 |
6 |
2 |
ACQ |
9 |
83 |
8 |
TTR |
1 |
11 |
88 |
Rows are what the emitter really was; columns are what the classifier said; the diagonal is where it was right. The call the aircrew cares about most is “TTR” — a radar that is tracking you — so treat it as the detection, exactly as L8 treated a target:
\(P_d\) is read along the row. Of the 100 true TTRs, 88 were called TTR: \(P_d = 88/100 = 0.88\). The other 12 are misses, and the 11 called “ACQ” are the dangerous ones — you are being tracked and your display says you are being searched for.
\(P_{fa}\) is read down the column, off the diagonal. Of the 200 emitters that were not TTRs, \(2 + 8 = 10\) were called TTR: \(P_{fa} = 10/200 = 0.05\). “Not a TTR” plays the role of L8’s noise-alone hypothesis.
Do not confuse either one with the fraction of TTR calls that were right (\(88/98\)). That number — the precision — depends on how common TTRs are, and it is worth a sentence of its own. Suppose only one intercept in twenty is really a TTR. In 1000 intercepts, the 50 TTRs produce \(0.88 \times 50 = 44\) correct calls, while the 950 other emitters produce \(0.05 \times 950 \approx 48\) false ones: about half of this classifier’s TTR warnings would be false. It is L8’s lesson about millions of range cells, arriving in a new place — when the thing you are looking for is rare, even a small \(P_{fa}\) fills the display.
One matrix is one point on a ROC#
A classifier rarely outputs a bare label; it outputs a score, and a threshold on that score decides when to say “TTR”. Lower the threshold and more true TTRs are caught — \(P_d\) rises — but more search radars are called TTR too — \(P_{fa}\) rises. Sweep the threshold and the pairs trace out exactly the receiver operating characteristic of L8; the matrix above is one operating point on it. Where to sit is the same kind of decision L8 called a commander’s preference: a missed TTR can cost the aircraft, while a false TTR costs a needless countermeasure or maneuver, so the threshold for this call is set low and the false alarms are accepted.
Where it breaks#
Out-of-library. The matrix has no unknown column. A classifier trained on three emitter types, shown a fourth, still answers with one of the three — often confidently. Unless it is built to reject inputs unlike anything it trained on, it turns the library’s “no match” into something worse: a confident, wrong match.
Adversarial inputs. An emitter designed by someone who understands the classifier can be shaped to be mislabeled — small, deliberate parameter changes that move it across a decision boundary.
The test is not the fight. A matrix measured on yesterday’s emitters says little about a reprogrammed one, which is where this section meets the last.
Key Concept
Read a confusion matrix the L8 way: pick the class that matters, read its row for \(P_d\) (the fraction of true members caught) and its column, off the diagonal, for \(P_{fa}\) (the fraction of non-members wrongly called it). Each matrix is one operating point on a ROC, and a classifier with no “unknown” answer will force a never-seen emitter into a known label.
The jammer’s reach: a denial circle#
The jammer’s side is Block 1’s free-space path loss. Its power at your receiver is its effective radiated power toward you, less the one-way path:
with ERP and \(J\) in dBm. At L1, \(20\log_{10} 1.57542 = 3.95\) dB. The ratio that matters is \(\text{J/S} = J - S\), and GPS is denied wherever it exceeds \((\text{J/S})_\text{tol}\). Because \(J\) depends only on range, that region is a circle around the jammer, and its radius is where J/S just reaches the threshold:
Work one. The companion script’s jammer has an ERP of 10 W, or 40 dBm. Then \(20\log_{10} R_\text{deny} = 40 + 130 - 30.1 - 3.95 - 92.45 = 43.5\) dB, and \(R_\text{deny} = 10^{43.5/20} \approx 150\) km. Ten watts — a modest transmitter — takes GPS away across a circle 300 km wide.
The reason is the geometry, not the jammer. The satellite’s signal pays for 20,000 km of path; the jammer pays for tens or hundreds. The jammer does not have to out-shout the satellite’s transmitter — only the satellite’s path. And the jammer’s path is one-way, which the type-along below turns into a scaling rule.
Two refinements matter for a route. A ground jammer can only reach what it can see, so the circle is capped by the radio horizon of L32, \(d = 4.12(\sqrt{h_r} + \sqrt{h_t})\) km with heights in meters: against an aircraft at 10 km altitude a 10 m mast sees out to about 425 km, so the circle binds first, but a low-flying route can hide below a jammer’s horizon. And the receiver’s antenna is not isotropic: a GPS antenna points its gain at the sky, and a jammer on the horizon sees less of it. An adaptive nulling antenna (the phased-array idea of L6, used for protection as in L15) pushes that much further, raising the J/S the receiver tolerates — the type-along asks what that is worth.
Type-along
One receiver, three jammers. Type this and read what it prints:
% L37 -- how far does a GPS jammer reach?
f = 1.57542; % GPS L1 [GHz]
S = -130; % received C/A power [dBm]
Gp = 10*log10(1.023e6/50); % chips per data bit -> G_p (L16)
M = 13; % margin the loops need after despread
JSt = Gp - M; % tolerable J/S at the antenna [dB]
PW = [1 4 16]; % jammer ERP toward us [W]
L = 10*log10(PW*1e3) - S - JSt; % path loss at J/S = JSt
R = 10.^((L - 20*log10(f) - 92.45)/20); % invert L2's FSPL [km]
fprintf('G_p = %.1f dB tolerable J/S = %.1f dB\n', Gp, JSt);
fprintf('%4.0f W -> GPS denied inside %5.1f km\n', [PW; R]);
Then answer these two:
Four times the power, twice the radius. Why only twice?
What would 6 dB more J/S tolerance do to the 16 W circle?
Solution
G_p = 43.1 dB tolerable J/S = 30.1 dB
1 W -> GPS denied inside 47.3 km
4 W -> GPS denied inside 94.6 km
16 W -> GPS denied inside 189.1 km
The jammer’s signal travels one way, so its power at the receiver falls as \(1/R^2\) and the range at which J/S reaches the threshold grows only as \(\sqrt{\text{ERP}}\). Four times the power is 6 dB, and 6 dB of path is one doubling of range (L2’s 6 dB rule): 47.3 km becomes 94.6 km becomes 189.1 km. A radar’s echo pays the path twice (\(1/R^4\)), so a radar needs sixteen times the power to double its range; the jammer needs four. That cheap reach is the whole NAVWAR problem.
Six more decibels of tolerance means the jammer must be 6 dB louder at the antenna, which it only is at half the range: the 16 W circle shrinks from 189.1 km to about 95 km — the same size as the 4 W jammer’s against the original receiver. Six decibels of receiver buys back exactly what six decibels of jammer took. That is why anti-jam antennas and more processing gain are the answers: each 6 dB halves the circle, and quarters the area it denies.
Quick problem — Inside the Circle?
A notional 2 W GPS jammer on a ridge; your receiver is 50 km from it.
GPS L1: 1575.42 MHz; received C/A power \(-130\) dBm
Tolerable J/S: 30 dB
Jammer ERP toward you: 2 W
Range to the jammer: 50 km
\(L_\text{fs} = 20\log_{10} R_\text{km} + 20\log_{10} f_\text{GHz} + 92.45\)
(a) The path loss, and the jammer power \(J\) at the receiver.
(b) The J/S, and whether the receiver keeps lock.
(c) The denial radius of this jammer.
Solution
(a) \(L_\text{fs} = 20\log_{10} 50 + 20\log_{10} 1.57542 + 92.45 = 33.98 + 3.95 + 92.45 =\) 130.4 dB. Two watts is \(10\log_{10} 2000 = 33.0\) dBm, so \(J = 33.0 - 130.4 =\) −97.4 dBm.
(b) \(\text{J/S} = -97.4 - (-130) =\) 32.6 dB — 2.6 dB above the tolerable 30 dB, so lock is lost: the receiver is inside the circle.
(c) GPS returns where the path costs 2.6 dB more, and 2.6 dB of one-way path is a factor of \(10^{2.63/20} = 1.354\) in range: \(R_\text{deny} = 50 \times 1.354 =\) 67.7 km. (Directly: \(20\log_{10} R_\text{deny} = 33.0 + 130 - 30 - 3.95 - 92.45 = 36.6\) dB, and \(10^{36.6/20} = 67.7\) km.)
So what: a two-watt transmitter holds GPS off a circle 135 km across, and backing out of it takes less than half a doubling of range — the 6 dB rule lets you estimate that in your head.
Jamming denies; spoofing deceives#
Everything so far has been jamming: noise that raises the floor until the tracking loops lose lock. It is a loud, honest failure. The receiver knows it has lost the signal, it can say so, and the aircraft can fall back on its inertial navigation system (INS) — the defenses being antennas that null the jammer, more processing gain, and an INS good enough to coast through the gap.
Spoofing is the other attack, and it is the harder one. A spoofer transmits counterfeit GPS signals, first matched to the real ones so the receiver’s loops lock onto them, then slowly walked away so the position solution follows. If you recognize the shape, you should: it is L18’s range-gate pull-off, aimed at your position instead of a radar’s tracking gate. The receiver may report no fault at all — just a confident, wrong position. The general defenses are authenticated signals (military codes are encrypted; civil signals are adding authentication), cross-checks against the INS, which drifts slowly and cannot be spoofed by radio, and integrity monitoring that flags a solution inconsistent with everything else the aircraft knows.
For a route planner the difference is stark. Jamming produces a known gap, and you can plan for it. Spoofing produces an unknown error, and a route flown on it can be flown precisely into the zone it was built to avoid.
The denied leg: INS drift meets the route#
Inside the denial circle the aircraft navigates on its INS, which integrates accelerations and rotations from the last good position. Its error grows with time. Real inertial error grows in a complicated way, but a notional constant drift rate \(\dot{e}\) (km/h) captures what a route planner needs:
where \(\ell\) is the length of the route inside the circle and \(V\) the ground speed. A route that passes a distance \(d\) from the jammer crosses a chord of length \(\ell = 2\sqrt{R_\text{deny}^2 - d^2}\). A navigation-grade INS drifts on the order of 1.5 km/h — the textbook “about a nautical mile per hour” class; lower-grade units drift an order of magnitude faster, and their error grows faster than linearly on a long leg.
The companion script flies exactly this picture. Its 10 W jammer denies a circle of 149.5 km (the script uses the derived threshold \(G_p - 13\) dB = 30.1 dB; a flat 30 dB gives 151 km); the route passes 60 km from the jammer, so the denied chord is \(2\sqrt{149.5^2 - 60^2} = 274\) km, and at 800 km/h the aircraft is without GPS for 20.5 minutes. A navigation-grade INS drifting at 1.5 km/h leaves the circle 0.51 km off; a notional lower-grade unit at 20 km/h leaves it 6.8 km off. Suppose the route was planned to pass 5 km outside a MEZ inside that circle: with the first INS the clearance is real, and with the second it is fiction — the aircraft may be inside the MEZ and not know it.
That is why the route planner must care. In L35’s language, a denied leg is a string of predict steps with no update: the covariance grows every step and nothing shrinks it. A planner that knows where the denial circles are can respond in three ways — inflate each MEZ by the error expected on that leg, prefer routes that spend less time denied, or plan a position fix before the leg that matters. What it cannot do is plan a 5 km clearance on a leg where the navigation is only good to 7 km.
MATLAB bundle · L37_GPSDenial.m↓
L37 has no interactive demo; the companion script L37_GPSDenial.m carries the lesson’s numbers. It plots J/S against range for the 10 W jammer with the tolerable level marked, tabulates the denial radius for three jammer powers beside the radio horizon, and plots the INS error across the denied chord for both drift rates against the 5 km clearance. Change the route’s closest approach d or the speed V and watch the exit error move.
Key Concept
GPS denial is a circle set by the jammer’s ERP, the one-way path loss, and the receiver’s tolerable J/S (\(G_p - M\)): its radius grows only as \(\sqrt{\text{ERP}}\), and every 6 dB of anti-jam halves it. Inside the circle the INS error grows as \(e \approx \dot{e}\,t\), so a route’s clearances are only as good as the navigation on that leg.
Space EW: uplink versus downlink#
Much of what supports a modern strike — communications, navigation, warning — passes through satellites, and satellite links can be jammed. The geometry decides who gets hurt.
Uplink jamming attacks the satellite. A simple relay (“bent-pipe”) satellite retransmits whatever arrives in its receive band from inside its receive beam. A jammer inside that beam is therefore relayed to every user in the downlink footprint — one jammer, a whole network. A jammer outside the beam is heard only through the antenna’s much weaker response in other directions, the sidelobe idea of L15, and barely registers.
Downlink jamming attacks the users. The satellite’s signal arrives at the ground faint after a long path, so a small jammer near a terminal wins easily — but only at the terminals near it. GPS denial is downlink jamming, which is exactly why it is a circle around the jammer rather than a global outage.
The protections are Block 2’s electronic protection moved to orbit: spot beams that narrow the region an uplink jammer must sit in, nulling receive antennas that steer a zero toward it, onboard processing that despreads the signal on the satellite so the jam is not relayed, and frequency hopping and spreading on the links themselves.
How four blocks set up these trends#
None of today’s trends needs new physics. Each one reuses what the course has already built, and changes who adapts faster:
Block |
What it gave you for today |
|---|---|
B1 · Radar fundamentals |
free-space path loss for the jammer’s reach (L2); detection theory and the ROC for reading a classifier (L8) |
B2 · RF EW |
ES classification and the threat library (L12); processing gain (L16) and J/S (L17) — the core of NAVWAR and space EW |
B3 · IR EW and signatures |
multi-band fusion (L28) and emission control (L27) — what an ML fuser combines and what a cognitive ES hunts |
B4 · Survivability and planning |
a route whose clearances (L32, L34) and covariance (L33, L35) must now absorb navigation error |
The Final Project: your Limitations section#
Section VI of your paper, Discussion and Limitations, must name at least one limitation drawn from this lesson. Pick one trend and say what it would change in your route — not in general, in yours:
Trend |
What it could change in your route |
|---|---|
NAVWAR |
the clearances past each MEZ must exceed the INS error on a denied leg; a denial circle becomes a cost layer |
Cognitive EW |
your detection contours come from library parameters; an agile radar’s may differ — widen the margins you rely on |
AI/ML emitter ID |
faster, surer identification raises each site’s \(P_{I\vert T}\), and with it every hazard \(\lambda_i\) on your route |
Space EW |
the SOJ timing, the SEAD window and pop-up cueing ride on links that may be denied — plan for support that arrives late |
A limitation earns its place when it changes a number or a line on your map. “GPS could be jammed” is a disclaimer. “A notional 10 W jammer near the target denies the last 270 km of my ingress; at 1.5 km/h my INS error there stays under 1 km, below my 4 km minimum MEZ clearance, so the route stands — but a lower-grade INS at 20 km/h would be about 7 km off by the end of that leg, inside the MEZ” is analysis. You can compute that one with this lesson’s equations and your own route.
Wrap-Up#
Four trends, four broken assumptions. Cognitive EW replaces the fixed threat library with a loop that senses, learns and adapts in the mission, because agile, software-defined emitters change faster than a library can be reprogrammed — and the adversary’s radars can run the same loop against your jamming. ML emitter identification is judged by a confusion matrix: a row gives \(P_d\), a column gives \(P_{fa}\), the matrix is one point on L8’s ROC, and a classifier with no “unknown” answer will confidently mislabel what it has never seen. NAVWAR starts from GPS’s \(-130\) dBm signal and the C/A code’s 43 dB of processing gain: a jammer denies a circle whose radius grows only as \(\sqrt{\text{ERP}}\), and inside it the INS error grows with every minute — so a route’s clearances are only as good as its navigation. Space EW turns on geometry: uplink jamming needs the satellite’s receive beam and reaches every user, downlink jamming is local. Next, L38 is the Final Project work day — the IEEE template, source vetting, and your figures. Bring one trend for your Limitations paragraph.