中國軍隊智慧化戰爭綜合觀察：聚焦智慧化戰爭中的反人工智慧作戰
現代英語：
Focus on anti-AI operations in intelligent warfare
■ Kang Ruizhi and Li Shengjie
introduction
The extensive application of science and technology in the military field has caused profound changes in the form and mode of warfare. The military game between major powers is increasingly manifested in technological subversion and counter-subversion, surprise and counter-surprise, offset and counter-offset. To win the future intelligent war, we must not only continue to promote the deep transformation and application of artificial intelligence technology in the military field, but also strengthen dialectical thinking, adhere to asymmetric thinking, innovate and develop anti-artificial intelligence combat theories and tactics, and proactively plan anti-artificial intelligence technology research and weapons and equipment research and development to achieve “breaking intelligence” and win, and strive to seize the initiative in future wars.
Fully understand the inevitability of anti-AI operations
Comrade Mao Zedong pointed out in “On Contradiction”: “The law of contradiction of things, that is, the law of the unity of opposites, is the most fundamental law of dialectical materialism.” Looking at the history of the development of military technology and its combat application, it has always been full of the dialectical relationship between attack and defense. The phenomenon of mutual game and alternating suppression between the “spear” of technology and the “shield” of corresponding counter-technology is common.
In the era of cold weapons, people not only invented eighteen kinds of weapons such as “knives, guns, swords, and halberds”, but also created corresponding “helmets, armor, and shields”. In the era of hot weapons, the use of gunpowder greatly increased the attack distance and lethality, but also gave rise to technical and tactical innovations represented by defensive fortifications such as “trench” and “bastion”. In the mechanized era, tanks shined in World War II, and people’s development of technical and tactical related to “tank armor” and “anti-tank weapons” continues to this day. In the information age, “electronic attack” and “electronic protection” around information control have set off a new wave of enthusiasm, and electronic countermeasures forces have emerged. In addition, there are countless opposing concepts in the military field such as “missiles” and “anti-missiles”, “unmanned combat” and “anti-unmanned combat”.
It should be noted that “anti-AI warfare”, as the opposing concept of “intelligent warfare”, will also gradually emerge with the extensive and in-depth application of intelligent technology in the military field. Prospective research on the concepts, principles and technical and tactical implementation paths of anti-AI warfare is not only a need of the times for a comprehensive and dialectical understanding of intelligent warfare, but also an inevitable move to seize the high ground of future military competition and implement asymmetric warfare.
Scientific analysis of anti-AI combat methods and paths
At present, artificial intelligence technology is undergoing a leapfrog development stage from weak to strong, and from special to general. From the perspective of its underlying support, data, algorithms, and computing power are still its three key elements. Among them, data is the basic raw material for training and optimizing models, algorithms determine the strategic mechanism of data processing and problem solving, and computing power provides hardware support for complex calculations. Seeking ways to “break intelligence” from the perspective of the three elements of data, algorithms, and computing power is an important method and path for implementing anti-artificial intelligence operations.
Anti-data operations. Data is the raw material for artificial intelligence to achieve learning and reasoning. The quality and diversity of data have an important impact on the accuracy and generalization ability of the model. There are many examples in life where artificial intelligence models fail due to minor data changes. For example, the face recognition model in the mobile phone may not be able to accurately identify the identity of the person because of wearing glasses, changing hairstyle or changes in the brightness of the environment; the autonomous driving model may also misjudge the road conditions due to factors such as road conditions, road signs and weather. The basic principle of implementing anti-data operations is to mislead the training and learning process or judgment process of the military intelligent model by creating “polluted” data or changing the distribution characteristics of the data, and use the “difference” of the data to cause the “error” of the model, thereby reducing the effectiveness of the military intelligent model. Since artificial intelligence models can conduct comprehensive analysis and cross-verification of multi-source data, anti-data operations should pay more attention to packaging false data information from multi-dimensional features to enhance its “authenticity”. In recent years, foreign militaries have conducted relevant experimental verifications in this regard. For example, special materials coating, infrared transmitting device camouflage and other methods are used to simulate the optical and infrared characteristics of real weapon platforms and even the engine vibration effects to deceive intelligent intelligence processing models; in cyberspace, traffic data camouflage is implemented to enhance the silent operation capability of network attacks and reduce the effectiveness of network attack detection models.
Anti-algorithm warfare. The essence of an algorithm is to describe a strategy mechanism for solving a problem in computer language. Since this strategy mechanism has a limited scope of adaptation, it may fail when faced with a wide variety of real-world problems. A typical example is Lee Sedol’s “God’s Move” in the 2016 man-machine Go match. After reviewing and analyzing the game, many professional Go players said that the “God’s Move” was not actually valid, but it worked for AlphaGo. Silva, the developer of AlphaGo, explained that Lee Sedol had hit an unknown loophole in the computer; there are also analyses that it may be that “this move” contradicts the Go logic of AlphaGo or is beyond its strategy learning range, making it unable to cope. The basic principle of implementing anti-algorithm warfare is to conduct logical attacks or logical deceptions against loopholes in the algorithm strategy mechanism and weaknesses in the model architecture to reduce the effectiveness of the algorithm. Anti-algorithm warfare should be combined with specific combat actions to achieve “misleading deception” against the algorithm. For example, drone swarm reconnaissance operations often use reinforcement learning algorithm models to plan reconnaissance routes. To address this situation, irregular or abnormal actions can be created to make the reward mechanism in the reinforcement learning algorithm model less effective or invalid, thereby achieving the goal of reducing its reconnaissance and search efficiency.
Anti-computing power operations. The strength of computing power represents the speed of converting data processing into information advantage and decision-making advantage. Unlike anti-data operations and anti-algorithm operations, which are mainly based on soft confrontation, the confrontation method of anti-computing power operations is a combination of soft and hard. Hard destruction mainly refers to the attack on the enemy’s computing power center, computing network facilities, etc., by cutting off its computing power to make it difficult for its artificial intelligence model to function; soft confrontation focuses on increasing the enemy’s computing power cost, mainly by creating a “fog” of war and data noise. For example, during combat, a large number of meaningless data such as images, audio, video, and electromagnetic are generated to contain and consume the enemy’s computing power resources, reducing the effective effect rate of its computing power. In addition, attacks can also be carried out on weak links in defense such as the support environment and supporting construction of computing power. The computing power center consumes huge amounts of electricity, and attacking and destroying its power support system can also achieve the effect of anti-computing power operations.
Proactively plan the construction of anti-AI combat capabilities
In any war, the right tactics are used to win. In the face of intelligent warfare, while continuing to promote and improve intelligent combat capabilities, it is also necessary to strengthen preparations for anti-AI operations, proactively plan theoretical innovations, supporting technology development, and equipment platform construction related to anti-AI operations, and ensure the establishment of an intelligent combat system that is both offensive and defensive, and integrated with defense and counterattack.
Strengthen the innovation of anti-AI combat theory. Scientific military theory is combat effectiveness. Whether it is military strategic innovation, military scientific and technological innovation, or other military innovations, they are inseparable from theoretical guidance. We must persist in emancipating our minds, broadening our horizons, strengthening dialectical thinking, and using the innovation of anti-AI combat theory as a supplement and breakthrough to build a theoretical system of intelligent combat that supports and serves to win the battle. We must insist on you fight yours and I fight mine, strengthen asymmetric thinking, and provide scientific theoretical support for seizing battlefield control through in-depth research on anti-AI combat concepts, strategies and tactics, and effectively play the leading role of military theory. We must persist in the integration of theory and technology, enhance scientific and technological cognition, innovation, and application, open up the closed loop between anti-AI combat theory and technology, let the two complement and support each other, and achieve deep integration and benign interaction between theory and technology.
Focus on the accumulation of anti-AI military technology. Science and technology are important foundations for generating and improving combat effectiveness. Once some technologies achieve breakthroughs, the impact will be subversive, and may even fundamentally change the traditional war offense and defense pattern. At present, major countries in the world regard artificial intelligence as a subversive technology and have elevated the development of military intelligence to a national strategy. At the same time, some countries are actively conducting research on technologies related to anti-AI operations and exploring methods of AI confrontation, with the intention of reducing the effectiveness of the opponent’s military intelligence system. To this end, we must explore and follow up, strengthen the tracking and research of cutting-edge technologies, actively discover, promote, and stimulate the development of technologies such as intelligent confrontation that have anti-subversive effects, seize the technological advantage at the beginning of anti-AI operations, and prevent enemy technological raids; we must also carefully select, focus on maintaining sufficient scientific rationality and accurate judgment, break through the technical “fog”, and avoid falling into the opponent’s technical trap.
Research and develop weapons and equipment for anti-AI operations. Designing weapons and equipment is designing future wars. What kind of wars will be fought in the future will determine what kind of weapons and equipment will be developed. Anti-AI operations are an important part of intelligent warfare, and anti-AI weapons and equipment will also play an important role on future battlefields. When developing anti-AI weapons and equipment, we must first keep close to battlefield needs. Closely combine combat opponents, combat tasks, and combat environments, strengthen anti-AI combat research, accurately describe anti-AI combat scenarios, and ensure that the demand for anti-AI combat weapons and equipment is scientific, accurate, and reasonable. Secondly, we must establish a cost mindset. The latest local war practices show that combat cost control is an important factor affecting the outcome of future wars. Anti-AI operations focus on interfering with and confusing the enemy’s military intelligence system. Increasing the development of decoy weapon platforms is an effective way to reduce costs and increase efficiency. By using low-cost simulations to show false targets to deceive the enemy’s intelligent reconnaissance system, the “brain-breaking” effect can be extended and amplified, and efforts can be made to consume its high-value strike weapons such as precision-guided missiles. Finally, we must focus on upgrading while building, using, and upgrading. Intelligent technology is developing rapidly and is updated and iterated quickly. We must closely track the opponent’s cutting-edge military intelligent technology applications, understand their intelligent model algorithm architecture, and continuously promote the application and upgrading of the latest anti-artificial intelligence technology in weapon platforms to ensure its efficient use on the battlefield.
現代國語：
關注智慧化戰爭中的反人工智慧作戰
■康睿智 李聖傑
引言
科學技術在軍事領域的廣泛運用，引起戰爭形態和作戰方式的深刻變化，大國軍事博弈越來越表現為技術上的顛覆與反顛覆、突襲與反突襲、抵消與反抵消。打贏未來智慧化戰爭，既要不斷推進人工智慧技術在軍事領域的深度轉化應用，還應加強辯證思維、堅持非對稱思想，創新發展反人工智慧作戰理論和戰法，前瞻佈局反人工智慧技術研究和武器裝備研發，實現「破智」制勝，努力把握未來戰爭主動權。
充分認識反人工智慧作戰必然性
毛澤東同志在《矛盾論》中指出：「事物的矛盾法則，即對立統一的法則，是唯物辯證法的最根本的法則。」縱觀軍事技術發展及其作戰運用歷史，從來都充滿了攻與防的辯證關系，技術之「矛」與相應反制技術之「盾」之間相互博弈、交替壓制的現象屢見不鮮壓制的現象屢見不鮮。
在冷兵器時代，人們不僅發明出「刀、槍、劍、戟」等十八般兵器，與之相應的「盔、甲、盾」等也被創造出來。熱兵器時代，火藥的使用大幅提升了攻擊距離和殺傷力，但同時也催生了以「塹壕」「棱堡」等防禦工事為代表的技術戰術創新。機械化時代，坦克在二戰中大放異彩，人們對「坦克裝甲」與「反戰車武器」相關技戰術的開發延續至今。資訊時代，圍繞制資訊權的「電子攻擊」與「電子防護」又掀起一陣新的熱潮，電子對抗部隊應運而生。此外，「導彈」與「反導」、「無人作戰」與「反無人作戰」等軍事領域的對立概念不勝枚舉。
應當看到，「反人工智慧作戰」作為「智慧化作戰」的對立概念，也必將隨著智慧技術在軍事領域的廣泛深度運用而逐漸顯現。前瞻性研究反人工智慧作戰相關概念、原則及其技戰術實現路徑，既是全面辯證認識智慧化戰爭的時代需要，也是搶佔未來軍事競爭高地、實施非對稱作戰的必然之舉。
科學分析反人工智慧作戰方法路徑
目前，人工智慧技術正經歷由弱向強、由專用向通用的跨越式發展階段。從其底層支撐來看，數據、演算法、算力依舊是其三大關鍵要素。其中，數據是訓練和優化模型的基礎原料，演算法決定了數據處理與問題解決的策略機制，算力則為復雜計算提供硬體支撐。從數據、演算法、算力三個要素的角度尋求「破智」之道，是實施反人工智慧作戰的重要方法路徑。
反數據作戰。數據是人工智慧實現學習和推理的原始素材，數據的品質和多樣性對模型的準確度和泛化能力有重要影響。生活中因為微小數據變化而導致人工智慧模型失效的例子比比皆是。例如，手機中的人臉識別模型，可能會因人戴上眼鏡、改變發型或環境明暗變化等原因，而無法準確識別身份；自動駕駛模型也會因路況、路標及天氣等因素，產生對道路情況的誤判。實施反數據作戰，其基本原理是通過製造“污染”數據或改變數據的分佈特徵，來誤導軍事智能模型的訓練學習過程或判斷過程，用數據之“差”引發模型之“謬”，從而降低軍事智能模型的有效性。由於人工智慧模型能夠對多源數據進行綜合分析、交叉印證，反數據作戰應更加註重從多維特徵出發，包裝虛假數據信息，提升其「真實性」。近年來，外軍在這方面已經有相關實驗驗證。例如，利用特殊材料塗裝、紅外線發射裝置偽裝等方式，模擬真實武器平台光學、紅外特徵甚至是發動機震動效果，用以欺騙智能情報處理模型；在網絡空間，實施流量數據偽裝，以提升網絡攻擊靜默運行能力，降低網絡攻擊檢測模型的效果。
反演算法作戰。演算法的本質，是用計算機語言描述解決問題的策略機制。由於這種策略機制的適應範圍有限，在面對千差萬別的現實問題時可能會失效，一個典型例子就是2016年人機圍棋大戰中李世石的「神之一手」。不少職業圍棋選手復盤分析後表示，「神之一手」其實並不成立，但卻對「阿爾法狗」發揮了作用。 「阿爾法狗」開發者席爾瓦對此的解釋是，李世石點中了電腦不為人知的漏洞；還有分析稱，可能是「這一手」與「阿爾法狗」的圍棋邏輯相悖或不在其策略學習範圍內，導致其無法應對。實施反演算法作戰，其基本原理是針對演算法策略機制漏洞和模型架構弱點，進行邏輯攻擊或邏輯欺騙，以降低演算法有效性。反演算法作戰應與具體作戰行動結合，達成針對演算法的「誤導欺騙」。例如，無人機群偵察行動常採用強化學習演算法模型規劃偵察路徑，針對此情況，可透過製造無規則行動或反常行動，致使強化學習演算法模型中的獎勵機制降效或失效，從而達成降低其偵察搜尋效率的目的。
反算力作戰。算力的強弱代表著將資料處理轉換為資訊優勢和決策優勢的速度。不同於反數據作戰和反演算法作戰以軟對抗為主，反算力作戰的對抗方式是軟硬結合的。硬摧毀主要指對敵算力中心、計算網絡設施等實施的打擊，通過斷其算力的方式使其人工智能模型難以發揮作用；軟對抗著眼加大敵算力成本，主要以製造戰爭“迷霧”和數據噪聲為主。例如，作戰時大批量產生影像、音訊、影片、電磁等多類型的無意義數據，對敵算力資源進行牽制消耗，降低其算力的有效作用率。此外，也可對算力的支撐環境和配套建設等防備薄弱環節實施攻擊，算力中心電能消耗巨大，對其電力支援系統進行攻擊和摧毀，也可達到反算力作戰的效果。
前瞻佈局反人工智慧作戰能力建構
凡戰者，以正合，以奇勝。面對智慧化戰爭，持續推進提升智慧化作戰能力的同時，也需強化對反人工智慧作戰的未雨綢繆，前瞻佈局反人工智慧作戰相關理論創新、配套技術發展和裝備平台建設，確保建立攻防兼備、防反一體的智慧化作戰體系。
加強反人工智慧作戰理論創新。科學的軍事理論就是戰鬥力，軍事戰略創新也好，軍事科技創新也好，其他方面軍事創新也好，都離不開理論指導。要堅持解放思想、開闊視野，強化辯證思維，以反人工智慧作戰理論創新為補充和突破，建構支撐和服務打贏制勝的智慧化作戰理論體系。要堅持你打你的、我打我的，強化非對稱思想，通過對反人工智慧作戰概念、策略戰法等問題的深化研究，為奪取戰場制智權提供科學理論支撐，切實發揮軍事理論的先導作用。要堅持理技融合，增強科技認知力、創新力、運用力，打通反人工智慧作戰理論與技術之間的閉環迴路，讓兩者互相補充、互為支撐，實現理論與技術的深度融合和良性互動。
注重反人工智慧軍事技術累積。科學技術是產生和提高戰鬥力的重要基礎，有些技術一旦取得突破，影響將是顛覆性的，甚至可能從根本上改變傳統的戰爭攻防格局。當前，世界各主要國家將人工智慧視為顛覆性技術，並將發展軍事智慧化上升為國家戰略。與此同時，也有國家積極進行反人工智慧作戰相關技術研究，探索人工智慧對抗方法，意圖降低對手軍事智慧系統效能。為此，既要探索跟進，加強對前沿技術的跟踪研究，積極發現、推動、催生智能對抗這類具有反顛覆作用的技術發展，在反人工智能作戰起步階段就搶佔技術先機，防敵技術突襲；還要精挑細選，注重保持足夠科學理性和準確判斷，破除技術“迷霧”，避免陷入對手技術陷阱。
研發反人工智慧作戰武器裝備。設計武器裝備就是設計未來戰爭，未來打什麼仗就發展什麼武器裝備。反人工智慧作戰是智慧化戰爭的重要組成部分，反人工智慧武器裝備也將在未來戰場上發揮重要作用。在研發反人工智慧作戰武器裝備時，首先要緊貼戰場需求。緊密結合作戰對手、作戰任務和作戰環境等，加強反人工智慧作戰研究，把反人工智慧作戰場景描述準確，確保反人工智慧作戰武器裝備需求論證科學、準確、合理。其次要樹立成本思維。最新局部戰爭實踐表明，作戰成本控制是影響未來戰爭勝負的重要因素。反人工智慧作戰重在對敵軍事智慧系統的干擾與迷惑，加大誘耗型武器平台研發是一種有效的降本增效方法。通過低成本模擬示假目標欺騙敵智能偵察系統，可將「破智」效應延伸放大，力爭消耗其精確制導導彈等高價值打擊武器。最後要注重邊建邊用邊升級。智慧技術發展速度快、更新迭代快，要緊密追蹤對手前沿軍事智慧技術應用，摸準其智慧模型演算法架構，不斷推動最新反人工智慧技術在武器平台中的運用升級，確保其戰場運用的高效能。