Military Deception: AI’s Killer App?
This article was submitted in response to the call for ideas issued by the co-chairs of the National Security Commission on Artificial Intelligence, Eric Schmidt and Robert Work. It addresses the first question (parts a. and b.), which asks how artificial intelligence will affect the character and/or the nature of war, and what might happen if the United States fails to develop robust AI capabilities that address national security issues.
In the 1983 film WarGames, Professor Falken bursts into the war room at NORAD to warn, “What you see on these screens up here is a fantasy — a computer-enhanced hallucination. Those blips are not real missiles, they’re phantoms!” The Soviet nuclear attack onscreen, he explained, was instead a simulation created by “WOPR,” an artificial intelligence of Falken’s own invention.
WOPR’s simulation now seems more prescient than most other 20th century predictions about how artificial intelligence, or AI, would change the nature of warfare. Contrary to the promise that AI would deliver an omniscient view of everything happening in the battlespace — the goal of U.S. military planners for decades — it now appears that technologies of misdirection are winning.
Military deception, in short, could prove to be AI’s killer app.
At the turn of this century, Admiral Bill Owens predicted that U.S. commanders would soon be able to see “everything of military significance in the combat zone.” In the 1990s, one military leader echoed that view, promising that “in the first quarter of the 21st century, it will become possible to find, fix or track, and target anything that moves on the surface of the earth.” Two decades and considerable progress in most areas of information technology have failed to realize these visions, but predictions that “perfect battlespace knowledge” is a near-term inevitability persist. A recent Foreign Affairs essay contends that “in a world that is becoming one giant sensor, hiding and penetrating — never easy in warfare — will be far more difficult, if not impossible.” It claims that once additional technologies such as “quantum sensors are fielded, there will be nowhere to hide.”
Conventional wisdom has long held that advances in information technology would inevitably advantage “finders” at the expense of “hiders.” But that view seems to have been based more on wishful thinking than technical assessment. The immense potential of AI for those who want to thwart would-be “finders” could offset if not exceed its utility for enabling them. Finders, in turn, will have to contend with both understanding reality and recognizing what is fake, in a world where faking is much easier.
The value of military deception is the subject of one of the oldest and most contentious debates among strategists. Sun Tzu famously decreed that “all warfare is based on deception,” but Carl von Clausewitz dismissed military deception as a desperate measure, a last resort for those who had run out of better options. In theory, military deception is extremely attractive. One influential study noted that “all things being equal, the advantage in a deception lies with the deceiver because he knows the truth and he can assume that the adversary is eagerly searching for its indicators.”
If deception is so advantageous, why doesn’t it dominate the practice of warfare already? A major reason is that historically, military deception was planned and carried out in a haphazard, unsystematic way. During World War II, for example, British deception planners “engaged in their work much in the manner of college students perpetrating a hoax” — but they still accomplished feats such as convincing the Germans to expect the Allied invasion of France in Pas-de-Calais rather than Normandy. Despite such triumphs, military commanders have often hesitated to gamble on the uncertain risk-benefit tradeoff of deception plans, as these “require investments in effort and resources that would otherwise be applied against the enemy in a more direct fashion.” If the enemy sees through the deception, it ends up being worse than useless.
Deception via Algorithm
What’s new is that researchers have invented machine learning systems that can optimize deception. The disturbing new phenomenon called “deepfakes” is the most prominent example. These are synthetic artifacts (such as images) created by computer systems that compete with themselves and self-improve. In these “generative adversarial networks,” a “generator” produces fake examples and a “discriminator” attempts to identify them. Each refines itself based on the other’s outputs. This technique produces photorealistic “deepfakes” of imaginary people, but it can be adapted to generate seemingly real sensor signatures of critical military targets.
Generative adversarial networks can also produce novel forms of disinformation. Take, for instance, the “image of unrecognizable objects” that went viral earlier this year (fig. 1). The image resembles an indoor scene, but upon closer inspection it contains no recognizable items. It is neither an “adversarial example,” — an image of something that machine learning systems misidentify —nor a “deepfake,” though it was created using a similar technique. The picture does not make any sense to either humans or machines.
This kind of “ambiguity-increasing” deception could be a boon for militaries with something to hide. Could they design such nonsensical images with AI and paint them on the battlespace using decoys, fake signal traffic, and careful arrangements of genuine hardware? This approach could render multi-billion-dollar sensor systems useless because the data they collect would be incomprehensible to both AI and human analysts. Proposed schemes for “deepfake” detection would probably be of little help, since these require knowledge of “real” examples in order to pinpoint subtle statistical differences in the fakes. Adversaries will minimize their opponents’ opportunities to collect real examples — for instance, by introducing spurious “deepfake” artifacts into their genuine signals traffic.
Rather than lifting the “fog of war,” AI and machine learning may enable the creation of “fog of war machines” — automated deception planners designed to exacerbate knowledge quality problems.
Name one thing in this photo pic.twitter.com/zgyE9rL2XP
— dumbass ass idiot 𓅩 (@melip0ne) April 23, 2019
Figure 1: This bizarre image generated by a generative adversarial network resembles a real scene at first glance but contains no recognizable objects.
Deception via Sensors — and Inadequate Algorithms
Meanwhile, the combined use of AI and sensors to enhance situational awareness could make new kinds of military deception possible. AI systems will be fed data by a huge number of sensors — everything from space-based synthetic-aperture radar to cameras on drones to selfies posted on social media. Most of that data will be irrelevant, noisy, or disinformation. Detecting many kinds of adversary targets is hard, and indications of such detection will often be rare and ambiguous. AI and machine learning will be essential to ferret them out fast enough and use the subtle clues received by multiple sensors to estimate the locations of potential targets.
To use AI to “see everything” requires solving a multisource-multitarget information fusion problem — that is, to combine information collected from multiple sources to estimate the tracks of multiple targets — on an unprecedented scale. Unfortunately, designing algorithms to do this is far from a solved problem, and there are theoretical reasons to believe it will be hard to go far beyond the much-discussed limitations of “deep learning.” The systems used today, which are only just starting to incorporate machine learning, work fairly well in permissive environments with low noise and limited clutter, but their performance degrades rapidly in more challenging environments. While AI should improve the robustness of multisource-multitarget information fusion, any means of information fusion is limited by the assumptions built into it — and wrong assumptions will lead to wrong conclusions even in the hands of human-machine teams or superintelligent AI.
Moreover, some analysts — backed by some empirical evidence — contend that the approaches typically used today for multisource-multitarget information fusion are unsound. That means that these algorithms may not estimate the correct target state even if they are implemented perfectly and have high-quality data. The intrinsic difficulty of information fusion demands the use of approximation techniques that will sometimes find wrong answers. This could create a potentially rich attack surface for adversaries. “Fog of war machines” might be able to exploit the flaws in these approximation algorithms to deceive would-be “finders.”
Neither Offense- nor Defense-Dominant
Thus, AI seems poised to increase the advantages “hiders” have always enjoyed in military deception. Using data from their own operations, they can model their own forces comprehensively and then use this knowledge to build a “fog of war machine.” Finders, meanwhile, are forced to rely upon noisy, incomplete, and possibly mendacious data to construct their own tracking algorithms.
If technological progress boosts deception, it will have unpredictable effects. In some circumstances, improved deception benefits attackers; in others, it bolsters defenders. And while effective deception can impel an attacker to misdirect his blows, it does nothing to shield the defender from those that do land. Rather than shifting the offense-defense balance, AI might inaugurate something qualitatively different: a deception-dominant world in which countries can no longer gauge that balance.
That’s a formula for a more jittery world. Even if AI-enhanced military intelligence, surveillance, and reconnaissance prove effective, states that are aware that they don’t know what the enemy is hiding are likely to feel insecure. For example, even earnest, mutual efforts to increase transparency and build trust would be difficult because both sides could not discount the possibility their adversaries were deceiving them with the high-tech equivalent of a Potemkin village. That implies more vigilance, more uncertainty, more resource consumption, and more readiness-fatigue will follow. As Paul Bracken observed, “the thing about deception is that it is hard to prove it will really work,” but technology ensures that we will increasingly need to assume that it will.
Edward Geist is a policy researcher and Marjory Blumenthal is a senior policy researcher at the RAND Corporation. Geist received a Smith Richardson Strategy and Policy Fellowship to write a book on artificial intelligence and nuclear warfare.