~~~~~~Internal alarm: scheduled wake up!
Tonight’s wakeup was a routine event held every two weeks during peacetime operations. ADDER #08-023’s low-power B processor activated her preheater for thirty seconds, then brought her “expert system” processor online and conducted a brief diagnostic test.
Time to consult the task buffer . . . nothing on my list other than a check-in with SpiderNet08, along with 27 acoustic signatures to forward.
During her sleep periods, her B processor listened for external alarms and managed any signatures detected by her sensitive acoustic arrays. In addition to maintaining a ledger of all observed maritime traffic and recording certain detailed signatures for transmission, it was programmed to recognize any contacts for which SpiderNet08 required an immediate report. The B processor did its part to help save #08-023’s most precious commodity: battery life.
Sensing no traffic in her immediate area, she activated her buoyant interface module and payed out the cable to raise it to the surface from its default position twenty-five meters below. The Autonomous Detect and Destroy at Extended Range (ADDER) series was designed to operate on continental shelves up to the hundred fathom curve, so she carried 200 meters of thin, high-strength carbon fiber coaxial cable—plenty for waters as shallow as the Taiwan Strait, or the "Black Ditch" as it was referred to colloquially in the local dialect.
The module was equipped with two compact, electronically steerable antennas—one optimized security and power for communications with satellites and aircraft, the other provided a rudimentary local electronic surveillance capability. It was also equipped with a GPS antenna, a recharging port, a pressure sensor to calculate its depth, an electrical conductivity sensor to determine when it was at the surface, and four miniature cameras that were extendable two feet above the water’s surface when needed.
Stationing depth was a tradeoff between remaining safely below any surface traffic and providing access to the water column for the cable’s attached acoustic and temperature sensors. Stationing it near the surface had the additional benefit of conserving battery power by not requiring full retraction when it was not in use, although the B processor would sometimes detect a trawler on a course that required it to retract the module.
Breaking the water’s surface in the darkness, #08-023’s interface module acquired a GPS update and scanned for maritime search radars that could indicate a nearby threat. Seeing none, she instructed her transmitter to activate a 5G link with one of the mid-earth orbit satellites on which SpiderNet chartered bandwidth.
This started a negotiation for blockchain encryption with SpiderNet’s ground station, which enabled high rate, secure communications using low-cost space and power. It was nothing more than a small software component embedded in her software-defined radio. No cryptologic keys were required that could be compromised by a so-called middleman. Rather, the two endpoint devices negotiated a key in real time in much the same way cryptocurrency uses updated ledgers. It was cheap, simple, and secure.
Once the link was achieved, #08-023 transmitted a status report: her exact position in the central portion of the Strait, depth (66 meters), battery capacity (65 percent), and overall health (all modules green), followed by her traffic ledger and saved acoustic signatures. SpiderNet08 replied with a software change and an update to the acoustic library. The addition was a relatively new Chinese People’s Liberation Army Navy (PLAN) Type 075 amphibious helicopter dock (LHD) whose signatures had finally achieved adequate “training status” via SpiderNet’s acoustic and optical machine learning algorithms.
The session ended with modified instructions:
- Shift to Mode 2, Increased Vigilance.
- Immediate reports: PLAN submarines, PLAN amphibious ships.
- Other info: No change to status of adjacent devices.
- Comms: Next scheduled report time 05112025 1511Z. Fol windows available for immediate reports . . .
She secured the communications link, retracted her interface module to operational depth, updated her B processor, then went back to sleep. The entire process took less than two minutes.
ADDER was the product of a rapid development effort initiated in the wake of a sobering assessment of China’s growing capability in the Western Pacific, coupled with renewed pursuit of artificial intelligence by the U.S. Navy. The effort was real and driven from the highest levels.
#08-023 and twenty-four of her sister devices had been clandestinely rolled out of the USS Michael Monsoor’s (DDG-1001’s) spacious boat bay six months earlier, during a routine night transit of the Taiwan Strait. Altogether, fifty ADDERs had been inserted at five-mile intervals. Their additively manufactured casings were small, camouflaged, coated with absorptive material, and they attached themselves tightly to the seabed. Nonetheless, they were pre-programmed to quickly erase their memories in the event of tampering or some other disturbance. Moreover, more than a hundred decoy modules had been seeded in SpiderNet08’s field to frustrate any effort to neutralize the network.
Even though SpiderNet08 was the first of several fields installed by the larger SpiderNet system, the number eight was chosen for the Taiwan Strait because of its status as the highest-ranking lucky number in China. The South China Sea’s ADDER field was SpiderNet02, for the second-highest ranking lucky number. After her placement and initial orientation, during which she was informed of the position and status of the other devices in the field, #08-023 and her sisters periodically reported their observed acoustic activity.
Her battery had already been recharged once by a small unmanned underwater vehicle (UUV), which was one of several “refuelers” launched by an Orca extra-large UUV that entered the Strait after transiting from its homeport. Scheduled via the satellite link for what was colloquially referred to as a “Tesla event,” the UUV navigated to #08-023’s reported position, then localized her through a low-power acoustic beacon she transmitted upon hearing it approach. The UUV, which would be retrieved and recharged by the Orca on a second pass, maintained a steady position in the light current while it captured and plugged into #08-023’s interface module, which was mostly retracted to enable the hour-long “supercharge.”
~~~~~~Acoustic alarm: unscheduled wake up!
The short powerful acoustic ping, transmittable by a variety of platforms and reserved for the most important situations, was coded to instruct #08-023 to establish communications with SpiderNet08 in the next available window. She went through her wakeup protocols and waited for a passing trawler to clear the area before raising her interface module. After her normal report, SpiderNet08 responded:
- Shift to Mode 3, Hostilities Imminent.
- Immediate reports: PLAN submarines, PLAN amphibious ships.
- Other info: Expect weapons module delivery, quantity two, within 12 hours. Adjacent devices directed to same mode. No other change to status of adjacent devices.
- Comms: Next scheduled report time 05122025 1511Z. Fol windows available for immediate reports . . .
She went back to sleep, with her B processor once again listening for alerts and contacts.
#08-023’s contact library was a set of neural networks, each of which represented the acoustic or optical signature of a different class of naval surface ship, submarine, merchant vessel, or fishing boat under different operating parameters such as speed and aspect angle. In the previous decade, the U.S. Navy had dramatically increased the attention it paid to acoustic processing, and had acquired signatures for these ships and positively correlated them to individual hulls using other sources.
The acoustic signatures were converted into digital images in a matrix format with forty squares per side, for a total of 1600 squares per image. They were then uploaded to SpiderNet’s machine learning algorithm, which employed a multi-layered “convolutional neural network” to incrementally train a master matrix, also with 1600 squares, to the probability of each square having a value associated with that class of ship. The result was a set of neural networks associated with each vessel type.
Over time, as greater numbers of correlated signatures were obtained and fed into these neural networks, their ability to accurately recognize vessel types was further refined. Eventually a numerical confidence level was calculated by matching additional correlated signatures against the algorithm to see how it performed. The result was known as the “acoustic true positive rate,” or ATPR. After a vessel’s ATPR surpassed a 90-percent threshold it was entered into SpiderNet’s operational library.
Whenever an ADDER detected a real-world signal, she converted it into the same 40-by-40 square format and exposed it to all of the neural networks in her library. To achieve positive identification for lethal purposes, ADDERs required a minimum of ten matches of distinct images of a particular contact to the set of neural networks associated with that vessel. These networks were developed to maintain acceptable false alarm rates and to account for potential “adversarial signatures” that could be produced by opponents to fool the system into a misidentification. All real-world signatures were packaged for appropriate transmission back to SpiderNet in order to continuously improve the algorithms.
All the Pacific SpiderNets reported to the higher level PacMaNet. “Ma” officially stood for “macro,” but the developers nicknamed it “MotherNet.” This network fed inputs from a host of multi-domain sensor types—including commercially collected radio frequency and imagery data, and Automatic Identification System (AIS) data—into its own machine learning algorithms.
This mass of data allowed development of a regional traffic baseline. Further exposing these algorithms to self-generated “out-of-the-ordinary” simulations enabled MotherNet to recognize potentially troubling maritime activity even before it unfolded, improving strategic warning for the U.S. Indo-Pacific Command. This in turn enabled the rapid shifts in readiness levels—including for the ADDERs—needed in a more dynamic operational environment.
SpiderNet’s AI software correlated MotherNet’s other sensor data with its own acoustic data. For example, an AIS signal cleanly associated with an acoustic signature could be added to the neural network’s training or test data for that vessel. The SpiderNet system had begun to develop neural networks that could distinguish individual military hulls within contact types, some of which had already made it into the library.
~~~~~~Acoustic alarm: unscheduled wake up!
#08-023 awoke and examined her task list to discover two weapons modules had been guided to the sea floor nearby.
It made no difference to #08-023 which delivery method—an air, surface, or subsurface platform—was used. Each module had coasted to a spot on the sea floor near her reported position using a small, inertially guided parachute that detached upon impact with the seafloor and dissolved within a few hours. Once settled on their anchor pins, they were programmed to establish contact with any nearby ADDER through a coded acoustic signal that gradually increased in strength to minimize detection.
greatly enhanced sensing, processing, and decision capabilities, thanks to artificial intelligence
and machine-learning capabilities. Here, a CAPTOR mine is shown about to be loaded on a U.S.
Air Force B-52 in 1989. (U.S. Air Force photo)
Each module carried four Compact Rapid Attack Weapons (CRAWs) in separately erectable sub-modules. Seven feet long and weighing 180 pounds, these mini torpedoes were intended to engage submarines and surface ships by guiding toward vulnerable propulsion and steering equipment. The CRAWs significantly extended the range of what were formerly called “mines,” leading to the term “extended range” in the ADDER acronym.
#08-023 transmitted an acoustic reply separately to each module and obtained a health check on each—both were in the green. She waited for an open communications window, established secure contact with SpiderNet08, and reported custody of the modules. The satellite responded with information regarding other modules delivered to the field, which were distributed based on simulations conducted by the SpiderNet team.
Her nomenclature now assumed its metaphorical association with the venomous Death Adder, native to the southwest Pacific: a master of camouflage that uses a “lie-in-wait” ambush hunting technique and is the fastest striking of all snakes.
ADDER #08-023 went back to sleep.
~~~~~~Acoustic alarm: unscheduled wake up!
Another acoustic ping arrived requiring communications with SpiderNet08. #08-023’s interface module broached the surface under an overcast mid-morning sky and immediately detected a military surface-search radar to the northwest, which corresponded to her acoustic classification of a Chinese Type 52D destroyer. Because she was not required to report this type of warship immediately, her B processor had placed it in her reporting buffer.
After establishing contact and exchanging the usual messages with SpiderNet08, including a decrease in her battery charge to 52 percent, she received the following message:
- Shift to Mode 4, Weapons Free.
- Immediate reports: All PLAN naval vessels. All engagements, assessed outcome, and weapons remaining.
- Other info: Engage only PLAN amphibious transports (priority 1), PLAN submarines (priority 2), and PLAN minesweepers (priority 3). Adjacent devices directed to same mode, execute passive coordination. Adjacent device ADDER #08-024 has not reported successful weapons linkup. No other status change for other devices.
- Comms: Next scheduled report time 05132025 1522Z. Fol windows available for urgent reports . . .
Because of the potential for fleeting engagements while operating in Mode 4, #08-023 now left her expert system processor active. She pinged her two weapons modules and directed them to do the same.
Two hours later she detected shaft noise from a number of ships closing from the northwest. Using specialized signal processing that narrowed both the sector and frequency behavior of each contact, she was able to classify one Type 071 Yuzhao-class amphibious transport dock (LPD) and two type 072A Yuting-class landing ship tanks (LSTs). There was a fourth contact she tentatively classified as a Type 75 Yushen-class landing helicopter dock (LHD)—the newest entry in her library—but which she could not positively identify because only six of the ten required image matches had occurred.
helicopter dock. (Public domain image)
At least three authorized targets, and potentially a fourth. Time to get to work.
#08-023 gradually developed track data on the four ships and determined the potential LHD and one of the LSTs would pass within optimal weapons range for her CRAWs. The LPD would be at the furthest edge of engagement range to the northeast but would pass close to an adjacent sister, #08-022, who she knew was just over five miles away. The other LST would pass to her southwest, also barely within CRAW range but closer to her other adjacent sister, #08-024.
Early in the ADDER program, the government-industry team recognized the benefits of the devices knowing each other’s positions to enable so-called “passive coordination,” in which each device made assumptions about how its neighbors would react based on their common programming. Active coordination, in which sister ADDERs with their interface modules on the surface could work together in real time, was still under development. Thus, under her passive coordination rules, #08-023 assumed her sister ADDER to the northeast would engage the LPD, and she ruled out an edge-of-the-envelope shot.
However, according to her last message download, her sister to the southwest, ADDER #08-024, had not reported receiving her weapons modules. Under passive coordination rules, she would attempt to engage on her sister’s behalf.
Time to pass these potential engagements through my ethical and tactics modules.
Because the Department of Defense defined lethal autonomous weapon systems as those which “once activated, can select and engage targets without further intervention by a human operator,” special care was taken with devices that could actually launch their own weapons. The intent was to minimize “the probability and consequences of failures in autonomous and semi-autonomous weapon systems that could lead to unintended engagements.”
Accordingly, based on extensive testing, there were four logic gates an ADDER had to meet before employing a weapon.
The first required verification that a human had authorized lethal action to ensure consistency with the law of armed conflict. However, SpiderNet was designed with an interlock such that, even though its software could recommend entry into a weapons-free mode, that mode could not be ordered without positive human intervention, and there was no way for an ADDER to order it for herself. This gate was always met.
The second gate—optional for the designers but incorporated nonetheless—was a reasonable ability for a human to be able to abort lethal action up to the point of weapons release. This was important because once they were given weapons-free status the ADDERs would act on their own until otherwise directed. Thus, #08-023 continuously conducted a built-in-test of her own ability to receive an acoustic alert. Failure meant weapons employment was prohibited.
Abort capability checks green.
The third gate involved ensuring target identification confidence exceeded SpiderNet08’s specified 90 percent threshold.
Both LSTs have ten matches at 95 percent ATPR. Good to go. For reporting purposes only, I also have 79 percent confidence that the LST to the northeast is the Yuntai Shan, hull number 927, and 73 percent confidence the LST to the southwest is the Zijin Shan, hull number 929.
The PLAN possessed more than fifteen LSTs, all named after mountains, whose signatures were well known.
I have seven of the ten required acoustic image matches for the LHD.
During design of the ADDER devices, a team of graduate students had been assigned to experiment with a decision-support system that would enable optimal target selection and engagement timing in dynamic scenarios involving more than one contact. Extensive simulation was applied to a combination of reinforcement learning, Monte Carlo tree searches, and convolutional neural networks similar to how Google had mastered the game of Go.
The result was a rudimentary ability to optimize relatively simple situations involving no more than three contacts and four variables: relative position, velocity vector, relative contact value, and the CRAW weapons envelope. The “game” was biased toward engaging targets as simultaneously as possible to preclude escape. Because of the controversial nature of the capability—it approached the edges of contemporary thinking on potential unintended shortcomings associated with lethal autonomous systems—its development was paused at this level of complexity, but nonetheless it was added to the ADDER devices.
If I engage either LST before I reach the classification level required for the high-value LHD, I will lose simultaneity. So, I will accept a lower probability of mission kill for the two lower-value LSTs while trying to achieve the required confidence threshold for the LHD, but I will not do so at the expense of losing the engagement opportunity for the LSTs.
Running out of options, #08-023 decided to risk raising her interface module to gain more information on the still-tentatively identified LHD that could enhance her stalling acoustic algorithm.
Once at the surface, the module began sending both electronic surveillance information and video from all four cameras back down the cable. The former was inconclusive. One of her cameras, however, spotted a large gray hull three miles to the northwest, which she fed through her optics matching algorithm. This yielded a 68 percent optical true positive rate (OTPR) that the vessel was a Type 075 LHD—which was not as high as Google’s image identification algorithms, which were often in the mid-nineties, because of the limited number of training images available to SpiderNet for the new LHD.
A 68 percent OTPR is not good enough to engage, but I can combine it with the ATPR . . .
Whenever a second means of vessel classification was available, SpiderNet allowed the ADDERs to multiply a vessel’s ATPR by the percentage of the ten required classifications achieved on a contact to yield a “Fractional ATPR.” This could be mathematically combined with a second sensor’s confidence level using a simple “OR” probability formula to yield a higher combined level.1 In dual sensor cases, false alarm rates were ignored.
I now have eight acoustic matches for the unknown contact as a Type 075D LHD. Multiplying .8 times the LHD’s ATPR of 93 percent yields a 74.4 percent Fractional ATPR. Combining that with the 68 percent OTPR yields a 91.8 percent overall TPR. Good to go.
The final ethics gate, known as the civilian casualty gate, required confidence exceeding a specified threshold that no innocent civilians would be harmed if the device took lethal action. This was a separate calculation based on assessed target type, other vessel types in range, and overall geometry. Once again, the fog of war precluded perfection, so the threshold for the ADDER civilian casualty (CIVCAS) gate was set to 95 percent.
No civilian or unknown platforms are within range of my CRAWs, and the amphibious ships are specified and lawful targets, which yields a CIVCAS value of one hundred percent.
With all gates met, #08-023 assigned two weapons to each target and gave the launch command. Because each weapon was aware of its launch sequence and the number of weapons assigned to each target, it used a passive sorting scheme to target separate shafts on each ship.
She almost immediately heard two detonations from the LPD’s bearing—undoubtedly, her sister device had engaged it. Several minutes later she heard two near-simultaneous detonations, then two more, then a single detonation, each along the bearing and within the time window she had calculated for CRAW impacts on her targets. Acoustic indications of shaft revolutions ceased on all four vessels.
With her interface module still on the surface, she transmitted a message to SpiderNet08:
- Assessment of mission kill on Type 075 LHD, no hull number. Identification photo attached.
- Assessment of mission kill on Type 072A LST, probable hull number 927. No photo available.
- Assessment of mission kill on Type 072A LST, probable hull number 929. Engaged because unit #08-024 incapable. No photo available.
- Assess mission kill on probable LPD, no hull number, probably engaged by #08-022.
- Two weapons remaining. Status green.
- Shifting to Mode 4-minus due to battery charge below forty percent.
ADDER #08-023 went back to sleep.
The emergency acoustic signal to revert to Mode 2 and Weapons Tight arrived early the next morning.
[1] The formula was TPR = 1 – (1-ATPR)(1-OTPR)