Attack as a Whole in the Cloud: Counting the Military Advantage Once Across Replicated Data Centres

attack as a whole
Photo by planet volumes on unsplash.com

Estimated reading time: 11 minutes


In post #33, I found that redundancy cuts both ways. It lowers the military advantage of striking one data centre and may lower the expected harm, but only where the attacker has information about it. Redundancy also gives an attacker a reason to strike several facilities that run the same service. On ratifying Additional Protocol I (AP I), several States declared that military advantage refers to the attack as a whole. How does that apply when the strikes hit several replicated data centres?

Two readings of the attack as a whole are in use. One weighs the whole attack against the whole harm. The other values each strike in light of the whole. For replicated services, the difference can decide the outcome. As I see it, the advantage of denying the service counts once. It is weighed against the expected harm of all the strikes together. Each strike still remains an attack with its own precautions.

I keep the cast of post #32 and post #33. States A and B are parties to an international armed conflict, and State B is the attacker. Provider P serves State A’s forces from several facilities in State A. I assume that each facility is a military objective. Civilians in State A and in neutral State N may still depend on the service. I do not assess any recent or ongoing operation.

I. The rules that apply

Attacks, proportionality and precautions

An attack is an act of violence against the adversary (Article 49(1) AP I). An attack is indiscriminate if its expected incidental harm would be excessive in relation to the concrete and direct military advantage anticipated (Article 51(5)(b) AP I). So is a bombardment that treats clearly separated objectives in a populated area as one (Article 51(5)(a)). Those who plan or decide upon an attack must take all feasible precautions in the choice of means and methods. They must refrain from excessive attacks and cancel or suspend an attack once excess becomes apparent (Article 57(2)(a)(ii)–(iii) and (b) AP I). Article 57(3) adds a choice rule. It applies where a choice is possible between several objectives offering a similar military advantage. The attacker must then choose the one whose attack may be expected to cause the least danger to civilians.

The proportionality rule is customary (Customary International Humanitarian Law Study of the International Committee of the Red Cross (ICRC), Rule 14), and I explained it in post #5. The customary status of Article 57(3) is less settled. The ICRC treats its counterpart as customary in international and arguably in non-international armed conflicts (Rule 21). A substantial majority of the Tallinn experts agree (Tallinn Manual 2.0, Rule 118, para. 1). The United States does not (US Department of Defense (DoD) Law of War Manual, § 5.11.7.1). Constant care, which governs military operations as a whole, is a separate duty (post #22).

The attack as a whole

Germany declared on ratifying AP I how it reads military advantage. It refers to the advantage anticipated from “the attack considered as a whole and not only from isolated or particular parts” (Germany, para. 5). The United Kingdom (SI 1998/1754, para. (j)) and the Netherlands (item 5) used the same phrase. The Netherlands applied this understanding expressly to Article 57(3) as well. The ICRC Study records such statements by several States (Rule 14, at p. 49).

The whole has limits. It is a finite operation and narrower than the war effort (International Law Association Study Group, at pp. 343 and 364). The DoD Manual reads the advantage more broadly. It may be assessed in “the full context of the war strategy” (§ 5.12.2.1). I return to that reading in Section III. Post #15 noted these declarations; here they carry the analysis.

Replicated services as an assumption

In this post, a service is replicated where it keeps running after the loss of any single facility. I treat this as a fact the attacker knows or does not know. I make no claim about how replication works. Where data is stored across several data centres, striking one offers a definite but limited advantage (Klonowska and Schmitt, “Targetable Objects Under IHL”). Milanovic and Schmitt go further. If redundancy is effective, disrupting military use of any one data centre may be too slight to offer a definite military advantage (“When Do Data Centers Become Military Objectives?”).

II. One attack, one balance

The same frame on both sides

Where the advantage is assessed for the attack as a whole, the expected harm must be assessed for the same whole. This is the premise of my analysis, and it is common ground. Gillard argues that both sides of the balance need the same frame of reference; anything else would undermine the balancing (Chatham House, 2018, para. 29). In her view, the wording of Article 51(5)(b) does not suggest otherwise (Gillard, Vanderbilt Journal of Transnational Law, 2018, at p. 829). At an expert meeting, the ICRC submitted that the same scale applies on both sides (ICRC, ed. Gisel, 2018, at p. 56). The DoD Manual requires the totality of the expected harm incidental to the attack to be considered (§ 5.12.1). These treatments converge, although that DoD passage does not address the attack as a whole.

When strikes on replicas form one attack

In my view, strikes on replicated facilities form one attack as a whole where the advantage sought is denial of the service. Gillard supplies the test. Attacks on different targets form one whole where the advantage of engaging one target is “dependent – in part or in full – on engaging other targets” (Gillard, Chatham House, 2018, para. 28). Where a strike’s advantage does not depend on and is not affected by other acts, that strike is the whole (para. 27). The parts of a whole may still be separate attacks, which must comply with the other rules on the conduct of hostilities (para. 26).

The idea is not new for networks. A committee established by the Prosecutor of the International Criminal Tribunal for the former Yugoslavia reasoned this way. It treated a strike on one part of a network as part of an integrated attack on many objects (Final Report, para. 78). According to the committee, the North Atlantic Treaty Organization had described that network as “hardened and redundant” (para. 72).

Replicated services fit this test. If the service keeps running after the loss of any single facility, the advantage of each strike depends on the others. The strikes then form one attack as a whole. The same dependence marks where the whole ends. A strike whose advantage does not depend on denying the replicated service falls outside it. Where a cyber operation accompanies kinetic strikes, the Tallinn experts assess it against the advantage of the entire attack (Tallinn Manual, Rule 113, para. 10).

III. Counting the advantage once

Two readings

The declarations of Germany, the United Kingdom and the Netherlands (Section I) fix the frame. They do not say how a single strike within it is weighed. Two readings are in use. The first weighs the whole attack against the whole harm. This is Gillard’s frame of reference, which Section II applied (Gillard, Chatham House, 2018, paras 27–29). The second reading was suggested at the ICRC expert meeting. There, the relevant advantage was said to be the one “offered by the specific attack, assessed in light of the attack as a whole” (ICRC, 2018, at p. 17). One expert added that gaining air supremacy by attacking all enemy aircraft and their support infrastructure is worth far more than attacking only part of them (at p. 17).

Other sources write in a similar way. The Tallinn experts assess the expected harm of a diversionary cyber attack in light of the advantage it contributed to the main attack (Tallinn Manual, Rule 113, para. 10). The DoD Manual notes that striking a communications relay yields more advantage as part of coordinated attacks on a command network than on its own (§ 5.12.2.1). None of these sources presents its reading as a rejection of the other. In practice, I expect the two readings to give the same result where each strike adds its own share of the advantage.

Why the advantage counts once

Replicated services are where the readings part. Where denying the service needs every strike, each strike is needed for the whole advantage. If each strike’s advantage is valued in light of the whole, each strike is credited with that full advantage. Each strike’s harm is then weighed against it. Across all the strikes, the attack can carry as much harm as if the advantage were counted once per strike.

In my assessment, this outcome is wrong. Article 51(5)(b) AP I relates the expected harm of an attack to the advantage of the same attack. Where the strikes form one attack, there is one advantage. It is counted once and weighed against the expected harm of all the strikes together. Denying a replicated service may justify more harm than a single strike would, because the advantage of the whole is larger. That larger advantage stands against the summed harm.

This question differs from the debate on cumulative harm. Gillard reports little support for weighing the harm of separate attacks cumulatively. She directs repeated high harm to the precautions instead (Chatham House, 2018, para. 30). Grouping within one attack as a whole is a different exercise. It applies only where the advantage itself depends on the group.

The opposing view

Some sources point to a broader frame. The DoD Manual states that military advantage may be assessed in the context of the war strategy (§ 5.12.2.1). It cites a 1991 US statement that proportionality is generally measured against an overall campaign (n. 446). The committee mentioned in Section II also said that proportionality need not focus only on a single incident. It referred to an overall assessment of civilian victims against the goals of the military campaign (Final Report, para. 78, citing para. 52). On that broader view, one could argue that limits per strike defeat the purpose of the declarations.

In my opinion, the broader reading does not change the counting problem. However wide the frame, the advantage of denying one service arises once. The frame also has limits. The International Law Association Study Group rejected a reading that measures advantage against the armed conflict as a whole, though on the definition of military objectives (at p. 343).

A limit within the whole?

The 1987 Commentary on Article 57 accepts that an attack carried out in a concerted manner in numerous places “can only be judged in its entirety” (para. 2218). It adds that this does not allow actions during such an attack that would cause severe civilian losses or extensive destruction. That limit matches the Commentary’s general view that incidental losses should never be extensive (Article 51, para. 1980). The majority of the Tallinn experts rejected that view. For them, extensive incidental harm may be lawful if the anticipated advantage is great enough (Tallinn Manual, Rule 113, para. 8). The limit in paragraph 2218 of the Commentary on Article 57 is therefore contested, and I do not rely on it. Counting the advantage once does much of its work: severe harm at one site stays in the sum and weighs against the single advantage.

IV. Precautions within the whole

Each strike remains an attack

Counting the advantage once does not merge the strikes into a single act. As Section II noted, each strike remains an attack (Article 49(1) AP I) and must comply with the other rules. The practitioner reading is that each strike carries its own duties under Article 57(2) AP I. These are to verify the objective, to take feasible precautions in the choice of means and methods, and to cancel or suspend. The Tallinn experts add a limit. An attacker need not choose alternative weapons or tactics that yield less military advantage (Tallinn Manual, Rule 116, para. 3). Those executing one strike may not know the advantage of the whole. That information is less likely to be apparent when an attack consists of several concerted strikes (Gillard, Chatham House, 2018, paras 160–161). The duty to cancel or suspend still applies once excess becomes apparent (Article 57(2)(b) AP I).

Article 57(3) between whole attacks

Article 57(3) AP I works differently once the advantage requires every replica. Within that attack, there is no choice among the replicas. Each is needed, so none can be exchanged for a less dangerous one. The rule then applies at the level of the attack as a whole. The Netherlands declared that military advantage in Article 57(3) refers to the attack as a whole (item 5). The Tallinn experts likewise determine the advantage for the choice of targets in light of the operation as a whole (Tallinn Manual, Rule 118, para. 7). The Tallinn commentary and the DoD Manual both state that the rule does not require the attacker to give up the advantage it seeks (Tallinn Manual, Rule 118, para. 7; DoD Manual, § 5.11.7).

In my reading, Article 57(3) therefore compares alternative attacks that offer a similar military advantage. A similar advantage means comparable military effects, not identical ones (Tallinn Manual, Rule 118, para. 6). The alternative must also be a reasonable option, more than a mere possibility (Tallinn Manual, Rule 118, para. 5). Where such an attack would cause less danger to civilians, Article 57(3) points to it. The 1987 Commentary gives examples of objectives chosen so that the same advantage is gained without endangering the population (paras 2227–2228).

Conclusion

The attack as a whole lets an attacker weigh the advantage of denying a replicated service. It also obliges the attacker to weigh the expected harm of all the strikes against that advantage, counted once. Valuing each strike in light of the whole would count the same advantage several times. Each strike stays an attack, with its own precautions. Article 57(3) compares alternative attacks that offer a similar advantage.

Where every replica is needed, Article 57(3) cannot point to the least dangerous one, because there is no choice among them. My conclusion is that none of this needs new treaty law. The existing rules handle replicated infrastructure once the advantage is counted once.

About the author

With more than 25 years of experience, Andreas Leupold is a lawyer trusted by German, European, US and UK clients.

He specializes in intellectual property (IP) and IT law and the law of armed conflict (LOAC). Andreas advises clients in the industrial and defense sectors on how to address the unique legal challenges posed by artificial intelligence and emerging technologies.

A recognized thought leader, he has edited and co-authored several handbooks on IT law and the legal dimensions of 3D printing/Additive Manufacturing, which he also examined in a landmark study for NATO/NSPA.

Connect with Andreas on LinkedIn