黃夢蘭 謝曉蘭 唐揚 袁天偉 陳超泉 呂世超 張衛(wèi)東 孫利民
摘 要:基于視頻監(jiān)控系統(tǒng)網(wǎng)絡(luò)化和智能化發(fā)展帶來的風(fēng)險,研究其隱蔽式網(wǎng)絡(luò)攻擊問題,目的在于調(diào)研大量隱蔽式網(wǎng)絡(luò)攻擊案例,總結(jié)針對視頻監(jiān)控系統(tǒng)的隱蔽式攻擊特異性。結(jié)合蜜罐技術(shù)在檢測網(wǎng)絡(luò)攻擊行為和發(fā)現(xiàn)攻擊線索等方面的獨特優(yōu)勢,梳理針對視頻監(jiān)控系統(tǒng)隱蔽式攻擊的蜜罐防御方法。針對監(jiān)控視頻蜜罐在視覺場景部署上的不足,介紹了一種深度場景偽造防御框架,將生成式AI大模型與視頻監(jiān)控蜜罐相結(jié)合。最后提出了面向視頻監(jiān)控系統(tǒng)的蜜罐防御技術(shù)的發(fā)展方向。
關(guān)鍵詞:監(jiān)控視頻系統(tǒng); 信息安全; 蜜罐; 隱蔽式攻擊; 生成式AI
中圖分類號:TP393.08 文獻標(biāo)志碼:A?文章編號:1001-3695(2024)05-003-1301-07
doi:10.19734/j.issn.1001-3695.2023.09.0409
Covert attacks and honeypot defense on video surveillance systems
Abstract:This paper investigated the issue of covert network attacks on video surveillance systems in light of their networking and intelligent development, which brought about new risks. The primary objective of this research was to explore numerous cases of covert network attacks and summarize the specific characteristics of such attacks targeting video surveillance systems. Additionally, this paper examined the unique advantages of honeypot technology in detecting network attack behaviors and identifying attack clues, and outlined the honeypot defense methods against covert attacks on video surveillance systems. Furthermore, considering the shortcomings in the visual scene deployment of video surveillance honeypots, this paper introduced a deep scene fabrication defense framework that combined generative AI models with video surveillance honeypots. Finally, it proposed the development direction of honeypot defense technology for video surveillance systems.
Key words:video surveillance system; information security; honeypot; covert attacks; generative AI
0 引言
視頻監(jiān)控系統(tǒng)作為一種安防設(shè)備,其應(yīng)用范圍遍布公共安防、城市交通、園區(qū)監(jiān)控、醫(yī)院、銀行、家庭等生產(chǎn)生活的各個場景。根據(jù)市場研究公司MarketsandMarkets[1]的數(shù)據(jù),預(yù)計視頻監(jiān)控市場將在2027年增長到764億美元,從2022—2027年,其復(fù)合年增長率(CAGR)將達到9.4%,這說明了各行各業(yè)對視頻監(jiān)控設(shè)備的需求日益劇增。視頻監(jiān)控系統(tǒng)可有效地監(jiān)測和記錄現(xiàn)實場景中的人員、車輛、物品等信息,為公共安全、商業(yè)保安、交通管理等領(lǐng)域提供強有力的技術(shù)支持。
然而,視頻監(jiān)控系統(tǒng)通常直接暴露在互聯(lián)網(wǎng)上,缺乏完善的安全防護措施,極易遭受網(wǎng)絡(luò)攻擊。據(jù)報道,約有73 000個安全攝像頭可使用默認密碼進行訪問[2],分布在全球256個國家,其中三星的閉路電視(closed-circuit television,CCTV)服務(wù)器至少有83 035臺暴露的設(shè)備,??低曋辽儆?60 415臺暴露的設(shè)備,只有53臺啟用了HTTPS安全防護措施。一些為利益或獵奇心理所驅(qū)使的黑客極有可能利用網(wǎng)絡(luò)空間搜索引擎FOFA[3]、Shodan[4]、Cydar[5]及鐘馗之眼(Zoomeye)[6]等工具探測發(fā)現(xiàn)暴露在互聯(lián)網(wǎng)上的設(shè)備和IP地址并對其實施攻擊。目前各個行業(yè)的視頻監(jiān)控系統(tǒng)相繼發(fā)生多起網(wǎng)絡(luò)安全事件,如入侵監(jiān)獄監(jiān)控系統(tǒng)[7]、威脅警察執(zhí)法[8]、盜竊銀行[9]和賭場[10]等資金轉(zhuǎn)運機構(gòu)涉密信息,影響人身安全[11],形成了一條集黑客破解、買賣、偷窺于一體的網(wǎng)絡(luò)黑產(chǎn)鏈。
與傳統(tǒng)的入侵檢測、防火墻等被動防御系統(tǒng)不同,蜜罐是一種用于發(fā)現(xiàn)網(wǎng)絡(luò)攻擊的主動防御系統(tǒng)。蜜罐能夠以低誤報、低干擾的方式部署運行,將攻擊者引入虛假的監(jiān)控系統(tǒng)中,在保護實際系統(tǒng)的同時捕獲攻擊數(shù)據(jù),并提供詳細和全面的攻擊信息以便對攻擊行為進行更深入的分析研究,實現(xiàn)更高效的攻擊預(yù)警和潛在的威脅發(fā)現(xiàn),因此在安全領(lǐng)域,蜜罐技術(shù)在攻擊線索識別方面具有明顯優(yōu)勢。對于未知威脅的研究,蜜罐技術(shù)已經(jīng)成為物聯(lián)網(wǎng)安全領(lǐng)域的重要技術(shù)[12],自2004年起,蜜網(wǎng)項目組團隊一直在推動蜜罐技術(shù)的發(fā)展,將蜜罐技術(shù)逐步融入到整個物聯(lián)網(wǎng)安全體系中[13]。蜜罐技術(shù)可以及時主動發(fā)現(xiàn)視頻監(jiān)控網(wǎng)絡(luò)攻擊線索,有效提升系統(tǒng)的主動防御能力。
本文通過回顧視頻監(jiān)控的發(fā)展歷程和趨勢,整理歸類針對視頻監(jiān)控系統(tǒng)的隱蔽式攻擊類型,并提出相應(yīng)的緩解措施;通過梳理物聯(lián)網(wǎng)蜜罐相關(guān)工作,重點介紹面向視頻監(jiān)控系統(tǒng)的蜜罐技術(shù)。
1 視頻監(jiān)控系統(tǒng)概述
1.1 視頻監(jiān)控系統(tǒng)架構(gòu)
視頻監(jiān)控系統(tǒng)通常由攝像頭、視頻錄像機、存儲設(shè)備、網(wǎng)絡(luò)設(shè)備等多個結(jié)構(gòu)組成[14,15]。其網(wǎng)絡(luò)拓撲結(jié)構(gòu)如圖1所示。這些組成部分通常是數(shù)據(jù)、軟件和基礎(chǔ)設(shè)施,承載著用戶的隱私和資產(chǎn)安全,因此它們是網(wǎng)絡(luò)空間攻擊的目標(biāo)。
攝像頭或網(wǎng)絡(luò)攝像頭負責(zé)捕捉監(jiān)控區(qū)域內(nèi)的視頻圖像,并將其轉(zhuǎn)換為數(shù)字信號。其中,視頻編碼和解碼器實現(xiàn)對數(shù)字信號的編碼和解碼,以便在網(wǎng)絡(luò)上傳輸和處理;網(wǎng)絡(luò)傳輸設(shè)備包括交換機、路由器等設(shè)備,用于承載和管理視頻數(shù)據(jù)的網(wǎng)絡(luò)傳輸;存儲設(shè)備包括數(shù)字視頻錄像機(digital video recorder,DVR)或網(wǎng)絡(luò)視頻錄像機(network video recorder,NVR)等,用于存儲從攝像頭獲取的視頻圖像;監(jiān)控終端設(shè)備包括顯示器、揚聲器、控制器等,用于顯示和控制監(jiān)控圖像;管理終端用于查看和管理視頻內(nèi)容的設(shè)備或應(yīng)用程序,如Web網(wǎng)站或者智能手機的應(yīng)用程序。
1.2 針對視頻監(jiān)控系統(tǒng)的網(wǎng)絡(luò)攻擊特異性
在攻擊目的方面,黑客選擇視頻監(jiān)控系統(tǒng)作為攻擊目標(biāo)的主要目的是竊取視頻數(shù)據(jù)、竄改監(jiān)控畫面或破壞監(jiān)控功能。視頻監(jiān)控系統(tǒng)本身可用于安防維護,其監(jiān)控記錄涉及的資產(chǎn)和隱私信息具有巨大的價值。通過獲取實時視頻流,攻擊者可以監(jiān)視特定區(qū)域,捕獲關(guān)鍵時刻的視頻和音頻信息,躲避身份識別,甚至隱藏犯罪行為、欺騙法律執(zhí)法機構(gòu)或扭曲真相。這種信息竊取多數(shù)以犯罪計劃、隱私侵犯、工業(yè)間諜活動或企業(yè)競爭為主要目的。
在攻擊手段方面,針對視頻監(jiān)控系統(tǒng)的攻擊更側(cè)重于利用監(jiān)控設(shè)施、通信協(xié)議和實時數(shù)據(jù)流來實施攻擊。由于視頻監(jiān)控系統(tǒng)的特殊性質(zhì)和功能,針對視頻監(jiān)控系統(tǒng)的攻擊可能會涉及物理破壞,比如損壞攝像頭或剪斷電纜。另外,視頻網(wǎng)絡(luò)監(jiān)控大多采用實時流媒體協(xié)議作為網(wǎng)絡(luò)傳輸協(xié)議[16~21],如RTSP、RTP、RTCP等傳輸協(xié)議。但是這些協(xié)議本身存在漏洞,攻擊者可能根據(jù)協(xié)議漏洞進行攻擊,如拒絕服務(wù)、中間人攻擊等。
針對視頻監(jiān)控系統(tǒng)的網(wǎng)絡(luò)攻擊與其他系統(tǒng)攻擊存在一定的區(qū)別,但網(wǎng)絡(luò)攻擊的一般原則和防御措施仍然適用于視頻監(jiān)控系統(tǒng),因此對于監(jiān)控系統(tǒng)的防御,除了及時更新軟件、加強訪問控制、使用加密傳輸、監(jiān)控網(wǎng)絡(luò)流量以檢測異常、進行安全培訓(xùn)等,還需要采取針對性的措施。
2 隱蔽式網(wǎng)絡(luò)攻擊
隱蔽式網(wǎng)絡(luò)攻擊是指攻擊者利用隱蔽手段和技術(shù),將攻擊和通信行為偽裝成合法的網(wǎng)絡(luò)流量和行為,躲避安全檢測和防護系統(tǒng)的發(fā)現(xiàn)和攔截,從而長期潛伏在目標(biāo)網(wǎng)絡(luò)中,以達到持續(xù)竊取數(shù)據(jù)和控制受害主機的目的。
2.1 隱蔽式攻擊特點和范疇
隱蔽式網(wǎng)絡(luò)攻擊最大的特點是其高度隱蔽性。攻擊者采用隱蔽的入侵技術(shù)躲避終端和網(wǎng)絡(luò)安全檢測,這使得其攻擊行為被安全軟件認為是合法程序或構(gòu)成部分,從而進行內(nèi)網(wǎng)移動,逐步入侵攻擊目標(biāo),并實現(xiàn)持續(xù)潛伏在目標(biāo)網(wǎng)絡(luò)中或合法網(wǎng)絡(luò)數(shù)據(jù)流中。
文獻[22]將隱蔽式網(wǎng)絡(luò)攻擊總結(jié)為四個特征:a)隱蔽性是對抗性入侵行為和攻擊流量的核心特征,攻擊者采用加密、偽裝等高級對抗技術(shù)實現(xiàn)隱蔽性,一些容易被異常檢測發(fā)現(xiàn)的網(wǎng)絡(luò)協(xié)議攻擊和違反IDS技術(shù)的攻擊不屬于隱蔽性攻擊的范疇;b)網(wǎng)絡(luò)相關(guān)性是指攻擊必須有網(wǎng)絡(luò)活動,即利用網(wǎng)絡(luò)作為控制信道和數(shù)據(jù)竊取信道,僅限于單機上的惡意軟件則不屬于隱蔽性攻擊的范疇;c)可控性是指攻擊者可以通過網(wǎng)絡(luò)遠程控制受害主機并執(zhí)行惡意操作,如記錄、下載或上傳指定數(shù)據(jù)等,傳統(tǒng)的蠕蟲和病毒因其可檢測性高、缺乏持久控制性、目的性單一,所以不屬于隱蔽性攻擊范疇;d)目的性是以持續(xù)竊取機密信息或長期控制利用為目的,因此普通的后門程序、常規(guī)的木馬和勒索軟件不屬于隱蔽性攻擊。例如,常規(guī)的拒絕服務(wù)(DoS/DDoS)攻擊就不屬于隱蔽式攻擊,其主要特點是利用大量的被感染或被控制的僵尸主機同時向目標(biāo)服務(wù)器發(fā)送大量請求,以消耗其資源,使其不可用。雖然在攻擊開始時可能存在一些隱蔽性,但DDoS攻擊的本質(zhì)是通過規(guī)?;恼埱髞碓斐善茐?,而不是通過隱蔽行為長時間存在于目標(biāo)系統(tǒng)內(nèi)。此外,Web入侵(如注入攻擊、跨站腳本注入等)、掃描攻擊、暴力破解等都因其主要目標(biāo)和執(zhí)行方式與隱蔽式攻擊的特征不相符而被劃分在隱蔽式攻擊范圍之外。
目前主流的網(wǎng)絡(luò)攻擊中,隱蔽型木馬和后門程序、新型僵尸網(wǎng)絡(luò)及高級持續(xù)性威脅(advanced persistent threat,APT)都屬于隱蔽式攻擊的范疇,然而并不是所有僵尸網(wǎng)絡(luò)和木馬都屬于此類攻擊范疇,因為早期的互聯(lián)網(wǎng)中繼聊天(Internet relay chat,IRC)僵尸網(wǎng)絡(luò)利用明文通信協(xié)議,通常具有明顯特征,常規(guī)木馬采用異常端口或明文協(xié)議,都容易檢測和識別,不符合高度隱蔽性特征。本文的隱蔽式攻擊主要是指類似APT的入侵或攻擊行為,但其所包含的范圍要廣于APT。
綜上所述,隱蔽式網(wǎng)絡(luò)攻擊并不是某個具體的入侵或攻擊類別,比如主流的惡意軟件功能復(fù)雜多樣,同時具備蠕蟲、后門、病毒、木馬等多種特征,難以準(zhǔn)確地將其歸類。本文更關(guān)注于具有高度隱蔽性、網(wǎng)絡(luò)控制性、目的性和持續(xù)控制這幾種共性技術(shù)或攻擊特征的攻擊。
2.2 針對視頻監(jiān)控系統(tǒng)的隱蔽式攻擊
雖然隱蔽式攻擊的一般原則仍然適用于視頻監(jiān)控系統(tǒng),但針對視頻監(jiān)控系統(tǒng)的攻擊更側(cè)重于利用監(jiān)控設(shè)備、通信協(xié)議和實時數(shù)據(jù)流來實施攻擊,其目的是通過竄改監(jiān)控畫面、竊取錄像、遠程控制攝像頭等手段,以破壞視頻監(jiān)控的真實性、完整性和可用性。本節(jié)調(diào)研了大量隱蔽式網(wǎng)絡(luò)攻擊案例,梳理出面向視頻監(jiān)控系統(tǒng)的隱蔽式攻擊,根據(jù)攻擊原理、目的、影響等特征對攻擊進行分類。表1中匯總了12種針對攝像頭的攻擊類型和案例。
這些攻擊大多采取措施來隱藏其痕跡,以便長期進行監(jiān)控、竊取敏感信息或操控監(jiān)控攝像頭。其中,代碼注入、中間人攻擊、偽造憑證、拒絕服務(wù)、社會工程學(xué)、供應(yīng)鏈風(fēng)險、逆向工程攻擊基本上都遵循一般隱蔽式攻擊的原則和技術(shù)。雖然拒絕服務(wù)攻擊本身并不隱蔽,但與一般的DoS攻擊形式相比(如SYN泛洪攻擊、Teardrop等),針對視頻監(jiān)控的DoS攻擊是在應(yīng)用層展開的。典型手段是打開并維護大量的實時流協(xié)議(real-time stream protocol,RTSP)連接,以此過度消耗網(wǎng)絡(luò)攝像頭資源,達到拒絕服務(wù)的目的[38,39]。除了通過泛洪攻擊來達到拒絕服務(wù)的目的外,攻擊者還可能偽裝成合法用戶向網(wǎng)絡(luò)攝像頭發(fā)送TearDown命令,惡意結(jié)束流媒體會話,使得攝像頭在真實客戶端不知情的情況下結(jié)束流媒體數(shù)據(jù)的傳輸。而中國作為全球人工智能監(jiān)控的主要驅(qū)動力,特別是華為、海康威視、大華和中興通信給63個國家提供了AI監(jiān)控技術(shù),針對機器學(xué)習(xí)算法的對抗性攻擊不容小覷。此外,視頻監(jiān)控大多數(shù)都連接在屏幕上,通常在攝像頭周圍安裝LED燈和麥克風(fēng),并且錄像視頻需要通過流媒體協(xié)議進行傳輸,因此產(chǎn)生了以太網(wǎng)挾持、流媒體協(xié)議漏洞利用和光學(xué)聲學(xué)通道攻擊問題。
2.3 蜜罐防御措施
傳統(tǒng)的網(wǎng)絡(luò)安全措施,如防火墻、認證控制、入侵預(yù)防系統(tǒng)(IPS)、入侵檢測系統(tǒng)(IDS)和基于行為的惡意軟件掃描儀等,通常依賴已知的威脅簽名而無法檢測新的攻擊形式,聚焦于攻擊的防御而忽視了攻擊的分析,往往產(chǎn)生大量的誤報,反而導(dǎo)致安全團隊浪費時間和資源;而且側(cè)重于保護網(wǎng)絡(luò)的邊界而忽視內(nèi)部網(wǎng)絡(luò)的安全,一旦被攻擊者突破,就可能任由攻擊者在內(nèi)部網(wǎng)絡(luò)自由行動[72]。即使采用深度防御策略[73],在整個目標(biāo)網(wǎng)絡(luò)中放置多層常規(guī)安全控制,網(wǎng)絡(luò)防御者仍然很難防止和檢測復(fù)雜的攻擊,如前面所提到的各種隱蔽式攻擊。這種有針對性的攻擊通常利用零日漏洞在目標(biāo)網(wǎng)絡(luò)上創(chuàng)建立足點,并且只留下少量惡意活動的痕跡以供檢測。
蜜罐[74]作為IPS和IDS相結(jié)合的主動防御工具,通過創(chuàng)建虛構(gòu)的賬戶和一些名字誘人的虛假文件以引誘黑客暴露攻擊行為和目標(biāo),誤導(dǎo)攻擊者浪費資源,延遲攻擊的效果。例如漏洞CVE-2017-11882可以通過Microsoft Office訪問并進一步獲取服務(wù)器的控制權(quán)。換句話說,欺騙性防御有助于建立一個積極的網(wǎng)絡(luò)防御態(tài)勢,其中的關(guān)鍵要素是在攻擊發(fā)生之前進行預(yù)測,增加攻擊的成本并收集新的威脅情報,以防止類似的攻擊,在檢測未知威脅方面展現(xiàn)出了巨大的優(yōu)勢。
目前,許多高交互蜜罐[75,76]能做到真實設(shè)備的仿真模擬,也有許多低交互蜜罐[77~80]滿足低成本部署的需求,混合蜜罐[81]的出現(xiàn)既滿足了蜜罐的高仿真度,也滿足了部署成本的降低。當(dāng)然,如果蜜罐是靜態(tài)的部署和配置,攻擊者會有足夠的時間識別蜜罐,進而終止攻擊。尤其是為入侵者提供真正的操作系統(tǒng)(OS)環(huán)境交互的高交互蜜罐,甚至可能被攻擊者利用來獲得控制權(quán)限,并用作跳板來損害其他系統(tǒng)。
隨著網(wǎng)絡(luò)攻擊越來越復(fù)雜,蜜罐技術(shù)也在不斷進步。對于蜜罐靜態(tài)部署和配置問題,Liu等人[82]基于擬態(tài)防御和移動目標(biāo)防御的思想設(shè)計了一種動態(tài)欺騙的隱蔽式攻擊防御系統(tǒng),其利用蜜罐吸引攻擊者和收集攻擊行為特征信息,基于SM4算法動態(tài)生成IP地址以提高蜜罐的可欺騙性,通過socket加密信道機制保證系統(tǒng)安全性;另外,根據(jù)蜜罐收集到的攻擊數(shù)據(jù),采用Viterbi算法訓(xùn)練和優(yōu)化模型,利用DHCPv6的動態(tài)策略分配機制定制策略并發(fā)送交互指令。Guarnizo等人[83]針對蜜罐靜態(tài)IP地址的局限性,提出了可擴展高交互蜜罐架構(gòu)SIPHON,該架構(gòu)在公共IP地址和IoT設(shè)備之間建立帶外信道惡意轉(zhuǎn)發(fā)通信,將實際地址在一個實驗室內(nèi)的設(shè)備描述為分布在世界各地的設(shè)備,實現(xiàn)了IP地址的偽造。該實驗僅通過7臺實驗室中真實的IoT設(shè)備,便虛擬出分布在世界不同地區(qū)的80多臺IP地址不同的高交互蜜罐。
此外,對于多階段的復(fù)雜攻擊,蜜罐攻防博弈中動態(tài)性地考慮根據(jù)收益與成本有效選擇蜜罐最佳防御策略。王鵑等人[84]利用軟件定義網(wǎng)絡(luò)(SDN)的高度靈活性、全局網(wǎng)絡(luò)可見性,提出了一種基于多階段攻擊響應(yīng)和動態(tài)博弈相結(jié)合的SDN動態(tài)蜜罐架構(gòu),通過攻擊圖將攻擊者的攻擊路徑劃分為不同階段并進行建模,利用不同交互類型蜜罐有針對性地進行防御響應(yīng);然后基于非合作不完全信息動態(tài)博弈模型構(gòu)建控制器決策引擎,對博弈雙方的攻防策略成本與收益進行量化分析,依據(jù)精煉貝葉斯均衡選取最優(yōu)蜜罐部署配置策略。與普通混合蜜罐系統(tǒng)和靜態(tài)蜜罐相比,SDN動態(tài)蜜罐在TCP連接速度上時間更長,其他指標(biāo)都明顯更高。
可擴展性是蜜罐部署中的一個痛點,雖然將蜜罐與移動目標(biāo)防御技術(shù)相結(jié)合滿足了動態(tài)IP地址配置,引入SDN技術(shù)和博弈理論解決了多階段攻擊的蜜罐部署,然而,這些蜜罐通常是針對某一款設(shè)備仿真模擬的,很難適用于各種供應(yīng)商的不同類型設(shè)備,也無法做到低成本地部署于世界各地。為了解決這個問題,Hakim等人[85]提出了一種針對UPnP協(xié)議的U-PoT高交互蜜罐框架,實現(xiàn)了忽略設(shè)備廠商、設(shè)備類型的可擴展性蜜罐部署,而且可以根據(jù)設(shè)備說明文件智能化創(chuàng)建蜜罐信息。UPnP技術(shù)利用IP、TCP、UDP、HTTP和XML等互聯(lián)網(wǎng)協(xié)議實現(xiàn)設(shè)備零配置并自動發(fā)現(xiàn)各種供應(yīng)商的不同設(shè)備類型。每個UPnP設(shè)備都是基于XML的設(shè)備模型定義的,因此U-PoT框架在可擴展的網(wǎng)絡(luò)環(huán)境中支持設(shè)備間的相互操作。
3 深度場景偽造的攝像頭蜜罐技術(shù)
物聯(lián)網(wǎng)蜜罐作為一種重要的安全防御工具,其技術(shù)已經(jīng)相對成熟并被廣泛應(yīng)用,如日志分析技術(shù)[86,87]、虛擬化技術(shù)[79,88]、重定向技術(shù)[89~91]等。但是大部分研究主要聚焦在監(jiān)控系統(tǒng)的物聯(lián)網(wǎng)架構(gòu)和通信風(fēng)險層面,比如設(shè)備認證、授權(quán)、訪問控制、安全通信、數(shù)據(jù)加密以及網(wǎng)絡(luò)態(tài)勢感知等手段,大多側(cè)重于整體系統(tǒng)的安全性。然而,針對攝像頭深度圖像偽裝的研究仍存在空白,攝像頭蜜罐的真實性和吸引力仍受限于虛擬的監(jiān)控場景和環(huán)境。
近年來,AI大模型技術(shù)熱度日益增高,成為當(dāng)今社會炙手可熱的話題。在ChatGPT發(fā)布后,與生成式AI大模型相關(guān)的arXiv論文數(shù)量急劇增加。最近的一篇論文調(diào)查表示,GPT-4[92]除了掌握語言外,在無須任何特殊提示的情況下可以解決數(shù)學(xué)、編碼、視覺、醫(yī)學(xué)、法律、心理學(xué)等領(lǐng)域復(fù)雜而困難的任務(wù),其處理問題的深度和廣度已經(jīng)接近人類級別,可視為通用人工智能(AGI)的早期版本。人工智能的快速發(fā)展推動了各個領(lǐng)域的研究范式發(fā)生革命性變化,在視覺生成領(lǐng)域,2021年OpenAI推出了CLIP[93]文生圖模型,從ResNet、Efficient、Vision Transformer到最大VIT large,這些計算機視覺領(lǐng)域的深度學(xué)習(xí)模型一步步探索融合,做到可以在無須任何打好標(biāo)簽的訓(xùn)練集的訓(xùn)練下,與之前完全有監(jiān)督方式訓(xùn)練出來的模型對比,準(zhǔn)確率相當(dāng)甚至更高。
本文順應(yīng)互聯(lián)網(wǎng)AI大模型發(fā)展的潮流,提出一種具備深度圖像偽裝能力的攝像頭蜜罐架構(gòu),利用AI大模型技術(shù)生成逼真的虛假監(jiān)控視頻,包括真實監(jiān)控系統(tǒng)中相似的場景、活動和人物,并將圖像偽裝部分搭建在攝像頭蜜罐系統(tǒng)上,極大地提高了蜜罐系統(tǒng)的真實性和吸引力。
3.1 深度視覺偽裝攝像頭蜜罐框架
面對攝像頭蜜罐場景部署的局限性,利用Diffusion[94]模型的圖像生成技術(shù)和智能補幀方法定向生成各類場景的監(jiān)控視頻,以提高攝像頭蜜罐的有效性。同時將偽裝視頻裝載到蜜罐所提供的Web平臺上,得到虛擬的視頻監(jiān)控管理系統(tǒng)來吸引潛在攻擊者。整體框架如圖2所示。
a)深度場景偽造模塊是框架的核心組成部分,其基于stable diffusion(SD)擴散模型,結(jié)合Temporal-kit和Ebsythn工具生成視頻關(guān)鍵幀,可自定義監(jiān)控場景描述,由該模塊生成與描述符合的偽裝畫面,得到逼真的視頻監(jiān)控場景。
b)Web攝像頭模塊負責(zé)接收來自深度圖像偽造模塊的偽裝場景視頻,并將其代替真實監(jiān)控視頻展示到Web服務(wù)器上。這一過程模擬了真實的監(jiān)控管理平臺,為用戶提供了與傳統(tǒng)監(jiān)控系統(tǒng)相似的用戶體驗。通過提供虛擬視頻流可以減少真實監(jiān)控系統(tǒng)的暴露風(fēng)險,同時保留了監(jiān)控操作的正常性,從而降低了潛在攻擊者的偵測難度。
c)蜜罐誘捕模塊負責(zé)提供Web攝像頭管理界面的HTTP服務(wù),并收集和分析潛在攻擊者的活動數(shù)據(jù)。該模塊能夠識別和記錄不正常的訪問行為并生成相關(guān)的報告或文檔,以幫助安全管理員及時發(fā)現(xiàn)和應(yīng)對潛在威脅。通過引誘攻擊者與虛擬監(jiān)控系統(tǒng)進行互動能夠更好地了解攻擊者的策略和技術(shù),從而提高系統(tǒng)的安全性和應(yīng)對能力。
3.2 OpenCanary開源蜜罐
在蜜罐誘捕模塊采用Python語言實現(xiàn)的OpenCanary開源密罐,因為代碼的開源,可以根據(jù)需求進行擴展和改寫,而且代碼結(jié)構(gòu)清晰,擴展簡單,設(shè)計不復(fù)雜,可以相對快速地實現(xiàn)想要的新功能。OpenCanary的基本實現(xiàn)原理是通過設(shè)置各種監(jiān)聽的端口模擬各種流行的服務(wù),而底層實現(xiàn)端口監(jiān)聽,依賴于Twisted的Python庫實現(xiàn)。圖3為OpenCanary蜜罐啟動HTTP服務(wù)并自定義Web界面的偽代碼。
OpenCanary蜜罐使用Twisted框架建立Web應(yīng)用服務(wù),代碼中定義了一個名為BasicLogin的類,它繼承自Resource類。這個類主要用于處理HTTP基本登錄認證。類中定義了兩個方法render_GET()和render_POST(),分別用于處理HTTP GET請求和HTTP POST請求。render_POST()嘗試從請求參數(shù)中獲取用戶名和密碼,并檢查是否匹配,如果匹配成功,則記錄日志并返回login.html頁面的內(nèi)容;否則返回錯誤頁面的內(nèi)容。render_GET()處理GET請求,如果沒有登錄失敗,則記錄日志并返回登錄頁面的內(nèi)容;否則返回錯誤頁面的內(nèi)容。本文的方法將默認打開的index.html文件替換成自定義的login.html文件,用于虛擬監(jiān)控系統(tǒng)的登錄界面,并在render_POST()函數(shù)中添加用戶名、密碼、頁面重定向的HTML文件。render_POST()獲取到登錄者匹配的用戶名和密碼后,將網(wǎng)頁重定向到模擬真實攝像頭的管理界面,用于展示各類監(jiān)控場景視頻。由于這些視頻是利用擴散模型和智能補幀方法所生成的偽造監(jiān)控視頻,管理界面所展示的監(jiān)控場景不涉及隱私信息。
4 未來展望
面向視頻監(jiān)控系統(tǒng)的蜜罐防御技術(shù)是非常重要的研究領(lǐng)域之一,可以幫助保護監(jiān)控系統(tǒng)免受隱蔽式網(wǎng)絡(luò)攻擊的威脅。目前,蜜罐防御技術(shù)已經(jīng)被廣泛研究和應(yīng)用,包括高交互蜜罐、低交互蜜罐、混合型蜜罐、虛擬蜜罐、實體蜜罐以及虛實結(jié)合的蜜罐等。機器學(xué)習(xí)在視頻監(jiān)控蜜罐領(lǐng)域中的應(yīng)用也是越來越廣泛,可以有效地對蜜罐數(shù)據(jù)進行分析和攻擊預(yù)測,提高蜜罐的誘捕效率和準(zhǔn)確性。面向視頻監(jiān)控的深度場景偽造密罐防御系統(tǒng)將繼續(xù)發(fā)展,未來的工作主要有以下幾點:
a)算法和模型進一步優(yōu)化。隨著深度學(xué)習(xí)和生成式AI模型的發(fā)展,算法和模型的進一步優(yōu)化可以保證系統(tǒng)能夠更準(zhǔn)確地模擬真實的監(jiān)控場景,提高對隱蔽式網(wǎng)絡(luò)攻擊的檢測和防御能力。
b)多模態(tài)數(shù)據(jù)融合。未來的深度場景偽造密罐防御系統(tǒng)可能會集成多種傳感器技術(shù)和數(shù)據(jù)源,如視頻、聲音、溫度等,以實現(xiàn)多模態(tài)數(shù)據(jù)的融合。這將增加系統(tǒng)對攻擊的感知能力,并提供更全面的安全保護。
c)實時動態(tài)場景偽造。目前的深度場景偽造密罐防御系統(tǒng)主要是基于離線生成的場景偽造。未來可以期待該系統(tǒng)具備實時動態(tài)場景偽造的能力,能夠根據(jù)環(huán)境變化自動生成逼真的監(jiān)控場景,提高系統(tǒng)的靈活性和適應(yīng)性。
利用AI大模型技術(shù)偽造監(jiān)控視頻仍然是一個相對新的研究方向。這一方向的研究與開發(fā)面臨諸多挑戰(zhàn),包括虛假視頻的逼真程度、制圖算法的準(zhǔn)確性、虛假視頻對攻擊者的吸引力等。未來面向視頻監(jiān)控系統(tǒng)的蜜罐防御技術(shù)將會朝著智能化的方向發(fā)展。
5 結(jié)束語
在視頻監(jiān)控系統(tǒng)的發(fā)展歷程中,隱蔽式網(wǎng)絡(luò)攻擊給監(jiān)控體系帶來了極大的威脅。針對視頻監(jiān)控系統(tǒng)網(wǎng)絡(luò)安全問題,蜜罐因其主動防御誘捕攻擊的特點得到廣泛關(guān)注,并成為一種有效的安全防御手段。本文查閱了大量隱蔽式網(wǎng)絡(luò)攻擊相關(guān)文獻,羅列出針對視頻監(jiān)控系統(tǒng)的隱蔽式攻擊,并提出蜜罐防御技術(shù)在視頻監(jiān)控系統(tǒng)安全防御的可靠性。本文結(jié)合攻防過程中遇到的蜜罐靜態(tài)地址和配置問題、蜜罐適配多階段攻擊問題和蜜罐部署擴展性問題,介紹了相關(guān)的防御技術(shù)。最后,本文緊跟互聯(lián)網(wǎng)發(fā)展趨勢,提出了基于擴散模型的深度場景偽造攝像頭蜜罐新方向,并展望其在網(wǎng)絡(luò)安全領(lǐng)域的潛在應(yīng)用和意義。
參考文獻:
[1]Video surveillance market by offering(camera, storage devices, monitors, AI-based VMS, non AI-based VMS, video content analysis, AI-driven video analytics, VSaaS), system (IP, analog, hybrid), resolution, vertical and region-global forecast to 2028, SE 2873[R]. Northbrook, IL: MarketsandMarkets, 2023.
[2]Smith. Peeping into 73000 unsecured security cameras thanks to default passwords[EB/OL]. (2014-11-06). https://www.csoonline.com/article/2844283/peeping-into-73-000-unsecured-security-cameras-thanks-to-default-passwords.html.
[3]FOFA search engine[EB/OL]. [2022-12-22]. https://fofa.info/.
[4]Simon K. Vulnerability analysis using Google and Shodan[C]//Proc of the 15th International Conference on Cryptology and Network Secu-rity. Cham: Springer, 2016: 725-730.
[5]物聯(lián)網(wǎng)識別發(fā)現(xiàn)系統(tǒng)[EB/OL]. [2022-12-22]. https://www.cydar.cn/. (IoT identification and discovery system[EB/OL]. [2022-12-22]. https://www.cydar.cn/.)
[6]ZoomEye[EB/OL]. [2022-12-22]. https://www.zoomeye.org/.
[7]Zhang Minlei, Chen Yancang, Chen Huan, et al. Design and implementation of a high performance network scanning system for VxWorks hosts[C]//Proc of International Conference on Communications, Information Management and Network Security. [S.l.]: Atlantis Press, 2016: 119-122.
[8]Dark Reading. Police car DVR P0wnage[EB/OL]. (2011-05-03). https://www.darkreading.com/cyber-risk/police-car-dvr-p0wnage.
[9]Aron J. Want to rob a bank? Hack your way in[J]. NewScientist, 2013, 220(2937): 5-65.
[10]WIRED. Crooks spy on casino card games with hacked security ca-meras, winEuclid Frakturd@p33M [EB/OL]. (2013-03-15). https://www.wired.com/2013/03/hackers-game-casino/.
[11]Kashmir H. How a creep hacked a baby monitor to say lewd things to a 2-year-old[EB/OL]. https://ww.forbes.com/sites/kosumirhill/2013/08/13/how-a-creep-hack-baly-monitor-to-say-lewd-thing-to-2-yer-old/?sh=2dze3243aad6.
[12]Metongnon L, Sadre R. Beyond telnet: prevalence of IoT protocols in telescope and honeypot measurements[EB/OL]. (2018-08-20). https://dl.acm.org/doi/pdf/10.1145/3229598.3229604.
[13]Zhang Weizhe, Qu Bin. Security architecture of the Internet of Things oriented to perceptual layer[J]. International Journal on Compu-ter, Consumer and Control, 2013,2(2): 37-45.
[14]吳群, 王田, 王漢武, 等. 現(xiàn)代智能視頻監(jiān)控研究綜述 [J]. 計算機應(yīng)用研究, 2016,33(6): 1601-1606. (Wu Qun, Wang Tian, Wang Hanwu, et al. Survey on modern intelligent video surveillance[J]. Application Research of Computers, 2016,33(6): 1601-1606.)
[15]Kalbo N, Mirsky Y, Shabtai A, et al. The security of IP-based video surveillance systems[J]. Sensors, 2020,20(17): 4806.
[16]曲柳鶯. 流媒體傳輸協(xié)議的研究[D]. 成都: 電子科技大學(xué), 2005. (Qu Liuying. Research on streaming media transmission protocol[D]. Chengdu: University of Electronic Science & Technology of China, 2005.)
[17]龔猷龍. 基于視頻監(jiān)控系統(tǒng)的RTP與SIP協(xié)議分析[D]. 北京: 北京郵電大學(xué),2008. (Gong Youlong. RTP and SIP protocols analysis based on surveillance system[D]. Beijing: Beijing University of Posts & Telecommunications, 2008.)
[18]Chu Dian, Jiang Chunhua, Hao Zongbo, et al. The design and implementation of video surveillance system based on H.264, SIP, RTP/RTCP and RTSP[C]//Proc of the 6th International Symposium on Computational Intelligence and Design. Piscataway, NJ: IEEE Press, 2013: 39-43.
[19]林型強. 多協(xié)議遠程視頻監(jiān)控的設(shè)計及基于Qt的實現(xiàn) [D]. 杭州:浙江工業(yè)大學(xué), 2020. (Lin Xingqiang. Design of multi-protocol remote video surveillance and implementation based on Qt[D].Hangzhou: Zhejiang University of Technology, 2020.)
[20]樊博. 基于SIP的網(wǎng)絡(luò)多媒體監(jiān)控系統(tǒng)設(shè)計[D]. 北京:北京理工大學(xué), 2016. (Fan Bo. Design of network multimedia monitoring system based on SIP [D]. Beijing: Beijing Institute of Technology, 2016.)
[21]Zhou Lin, Chen Yingmei, Li Zhen, et al. An improved video monitoring system based on RSVP protocol[C]//Proc of International Conference on Intelligent Transportation, Big Data and Smart City. Piscataway,NJ: IEEE Press, 2015: 94-97.
[22]曹自剛. 隱蔽式網(wǎng)絡(luò)攻擊檢測關(guān)鍵問題研究[D]. 北京: 北京郵電大學(xué), 2015. (Cao Zigang. Research on the key problems of covert network attack detection[D]. Beijing: Beijing University of Posts & Telecommunications, 2015.)
[23]Constantin L. Widely used wireless IP cameras open to hijacking over the Internet, researchers say[EB/OL]. (2013-04-11). https://www.pcworld.com/article/451274/widely-used-wireless-ip-cameras-open-to-hijacking-over-the-internet-researchers-say.html.
[24]Davies C. Google glass exploit hacked wearable with QR codes[EB/OL]. (2013-07-17). https://www.slashgear.com/google-glass-exploit-hacked-wearable-with-qr-codes-17290717.
[25]Castiglione A, Cepparulo M, De Santis A, et al. Towards a lawfully secure and privacy preserving video surveillance system[C]//Proc of the 11th International Conference on Electronic Commerce and Web Technologies. Berlin: Springer-Verlag, 2010: 73-84.
[26]Costin A. Security of CCTV and video surveillance systems: threats, vulnerabilities, attacks, and mitigations[C]//Proc of the 6th International Workshop on Trustworthy Embedded Devices. New York: ACM Press, 2016: 45-54.
[27]Dos Santos D. Sabotaging common IoT devices in smart buildings by exploiting unencrypted protocols[EB/OL]. (2019-07-30). https://www.forescout.com/blog/sabotaging-smart-building-iot-devices-using-unencrypted-protocols/.
[28]Li Hao, Xiezhang Tianhao, Yang Cheng, et al. Secure video surveillance framework in smart city[J]. Sensors, 2021,21(13): 4419.
[29]汪周紅. 面向視頻監(jiān)控網(wǎng)絡(luò)的入侵檢測技術(shù)研究[D]. 南京:東南大學(xué), 2021. (Wang Zhouhong. Research on intrusion detection technology for video surveillance network[D]. Nanjing: Southeast University, 2021.)
[30]Dowling S, Schukat M, Melvin H. A ZigBee honeypot to assess IoT cyberattack behaviour[C]//Proc of the 28th Irish Signals and Systems Conference. Piscataway, NJ: IEEE Press, 2017: 1-6.
[31]Dima B I. Breaking down Mirai: an IoT DDoS botnet analysis[EB/OL]. (2016-10-26). https://www.imperva.com/blog/malware-analysis-mirai-ddos-botnet/.
[32]Li Chao, Jiang Wei, Zou Xin. Botnet: survey and case study[C]//Proc of the 4th International Conference on Innovative Computing, Information and Control. Piscataway, NJ: IEEE Press, 2009: 1184-1187.
[33]Antonakakis M, April T, Bailey M, et al. Understanding the Mirai botnet[C]//Proc of the 26th USENIX Conference on Security Symposium. Berkeley, CA: USENIX Association, 2017: 1093-1110.
[34]Rashwan H A,Solanas A,Puig D,et al. Understanding trust in privacy-aware video surveillance systems[J]. International Journal of Information Security, 2016,15(6): 225-234.
[35]Kim J, Lee D, Park N. CCTV-RFID enabled multifactor authentication model for secure differential level video access control[J]. Multimedia Tools and Applications, 2020,79(8): 23461-23481.
[36]騰訊安全玄武實驗室. Looping surveillance cameras through live editing of network streams[EB/OL]. (2015-12-25). https://sec.today/events/talk/8de55eca-095b-4585-8771-d2e5ae68ae8a/. (Tencent Security Xuanwu Lab. Looping surveillance cameras through live editing of network streams[EB/OL]. (2015-12-25). https://sec.today/events/talk/8de55eca-095b-4585-8771-d2e5ae68ae8a/.)
[37]Huang Yong, Li Xiang, Wang Wei, et al. Forgery attack detection in surveillance video streams using Wi-Fi channel state information[J]. IEEE Trans on Wireless Communications, 2022,21(6): 4340-4349.
[38]Patil A, Banerjee S, Borkar G. A survey on securing smart gadgets using lightweight cryptography[C]//Proc of International Conference on Wireless Communication. Singapore: Springer, 2020: 503-515.
[39]褚書涵. 流媒體服務(wù)攻擊防御方案的設(shè)計[D]. 上海: 復(fù)旦大學(xué), 2011. (Chu Shuhan. Design of attack defense scheme for streaming media services[D]. Shanghai: Fudan University, 2011.)
[40]Kumar P A R, Selvakumar S. Distributed denial of service attack detection using an ensemble of neural classifier[J]. Computer Communications, 2011, 34(11): 1328-1341.
[41]Kumar P A R, Selvakumar S. Detection of distributed denial of ser-vice attacks using an ensemble of adaptive and hybrid neuro-fuzzy systems[J]. Computer Communications, 2013,36(3): 303-319.
[42]Zargar S T, Joshi J, Tipper D. A survey of defense mechanisms against distributed denial of service(DDoS) flooding attacks[J]. IEEE Communications Surveys & Tutorials, 2013,15(4): 2046-2069.
[43]Saied A, Overill R E, Radzik T. Detection of known and unknown DDoS attacks using artificial neural networks[J]. Neurocomputing, 2016,172(1): 385-393.
[44]VideoJak. What is VideoJak[EB/OL].[2022-12-22]. https://videojak.sourceforge.net/index.html.
[45]Salahdine F, Kaabouch N. Social engineering attacks: a survey[J]. Future Internet, 2019, 11(4): 89.
[46]Krombholz K, Hobel H, Huber M, et al. Advanced social enginee-ring attacks[J]. Journal of Information Security and applications, 2015, 22(6): 113-122.
[47]Pavkovic' N, Perkov L. Social engineering toolkit—a systematic approach to social engineering[C]//Proc of the 34th International Convention MIPRO. Piscataway, NJ: IEEE Press, 2011: 1485-1489.
[48]Ilascu I. Another supply chain mystery: IP cameras ship with malicious software[EB/OL]. (2016-04-14).https://securityledger.com/2016/04/another-supply-chain-mystery-ip-cameras-ship-with-malicious-software/.
[49]Paganini P. Be careful products sold on Amazon are infected with malware[EB/OL]. (2016-04-10). https://securityaffairs.com/46170/malware/amazon-products-infected-with-malware.html.
[50]Roberts P. Supply chain wreck: CCTV firmware vulnerable[EB/OL]. (2016-03-24). https://securityledger.com/2016/03/supply-chain-wreck-cctv-firmware-vulnerable/.
[51]Tekeoglu A, Tosun A S. Investigating security and privacy of a cloud-based wireless IP camera: NetCam[C]//Proc of the 24th Internatio-nal Conference on Computer Communication and Networks. Pisca-taway,NJ:IEEE Press,2015:1-6.
[52]SEC Consult. Backdoor in Sony IPELA engine IP cameras[EB/OL].(2016-12-06).https://sec-consult.com/blog/detail/backdoor-in-sony-ipela-engine-ip-cameras/.
[53]Exploiting network surveillance cameras like a Hollywood Hacker[EB/OL]. (2013-08-01). https://infocondb.org/con/black-hat/black-hat-usa-2013/exploiting-network-surveillance-cameras-like-a-hollywood-hacker.
[54]Heffner C. Exploiting surveillance cameras[EB/OL]. (2013-02-25). https://media.blackhat.com/us-13/US-13-Heffner-Exploiting-Network-Surveillance-Cameras-Like-A-Hollywood-Hacker-WP.pdf.
[55]Kim M J, Lee J Y, Chang H Y, et al. Design and performance evaluation of binary code packing for protecting embedded software against reverse engineering[C]//Proc of the 13th IEEE International Symposium on Object/Component/Service-Oriented Real-Time Distributed Computing. Piscataway, NJ: IEEE Press, 2010: 80-86.
[56]Lim K, Jeong J, Cho S, et al. An anti-reverse engineering technique using native code and obfuscator-LLVM for android applications[C]//Proc of International Conference on Research in Adaptive and Convergent Systems. New York: ACM Press, 2017: 217-221.
[57]Bursztein E. Attacks against machine learning: an overview[EB/OL]. (2018). https://elie.net/blog/ai/attacks-against-machine-learning-an-overview/.
[58]Feldstein S. The global expansion of AI surveillance[M]. Washington, DC: Carnegie Endowment for International Peace, 2019.
[59]Kadim Z, Johari K M, Samaon D F, et al. Real-time deep-learning based traffic volume count for high-traffic urban arterial roads[C]//Proc of the 10th IEEE Symposium on Computer Applications & Industrial Electronics. Piscataway, NJ: IEEE Press, 2020: 53-58.
[60]Chung J, Sohn K. Image-based learning to measure traffic density using a deep convolutional neural network[J]. IEEE Trans on Intelligent Transportation Systems, 2018,19(5): 1670-1675.
[61]Zhang Guohui, Wang Yinhai. Machine learning and computer vision-enabled traffic sensing data analysis and quality enhancement[M]// Data-Driven Solutions to Transportation Problems.[S.l.]: Elsevier Inc., 2019: 51-79.
[62]Tramèr F, Zhang Fan, Juels A, et al. Stealing machine learning models via prediction APIs[C]//Proc of the 25th USENIX Confe-rence on Security Symposium. Berkeley, CA: USENIX Association, 2016: 601-618.
[63]Pan J. Physical integrity attack detection of surveillance camera with deep learning based video frame interpolation[C]//Proc of IEEE International Conference on Internet of Things and Intelligence System. Piscataway, NJ: IEEE Press, 2019: 79-85.
[64]Mundra K, Modpur R, Chattopadhyay A, et al. Adversarial image detection in cyber-physical systems[C]//Proc of the 1st ACM Workshop on Autonomous and Intelligent Mobile Systems. New York: ACM Press, 2020: 1-5.
[65]Papernot N, Song Shuang, Mironov I, et al. Scalable private learning with pate[EB/OL]. (2018-02-24). https://arxiv.org/pdf/1802.08908.pdf.
[66]Papernot N, Abadi M, Erlingsson U, et al. Semi-supervised know-ledge transfer for deep learning from private training data[EB/OL]. (2017-03-03). https://arxiv.org/pdf/1610.05755.pdf.
[67]Brocker M, Checkoway S. iSeeYou: disabling the MacBook webcam indicator LED[C]//Proc of the 23rd USENIX Security Symposium. Berkeley, CA: USENIX Association, 2014: 337-352.
[68]Guri M, Zadov B, Elovici Y. LED-it-GO: leaking (a lot of) data from air-gapped computers via the(small) hard drive LED[C]//Proc of the 14th International Conference on Detection of Intrusions and Malware, and Vulnerability Assessment. Cham: Springer, 2017:161-184.
[69]Guri M, Hasson O, Kedma G, et al. VisiSploit: an optical covert-channel to leak data through an air-gap[EB/OL]. (2016-07-19). https://arxiv.org/ftp/arxiv/papers/1607/1607.03946.pdf.
[70]Cui Ang, Stolfo S J. A quantitative analysis of the insecurity of embedded network devices: results of a wide-area scan[C]//Proc of the 26th Annual Computer Security Applications Conference. New York: ACM Press, 2010: 97-106.
[71]Kanter G S, Reilly D, Smith N. Practical physical-layer encryption: the marriage of optical noise with traditional cryptography[J]. IEEE Communications Magazine, 2009,47(11): 74-81.
[72]蹇詩婕,盧志剛,杜丹, 等. 網(wǎng)絡(luò)入侵檢測技術(shù)綜述[J]. 信息安全學(xué)報, 2020,5(4): 96-122. (Jian Shijie, Lu Zhigang, Du Dan, et al. Overview of network intrusion detection technology[J]. Journal of Cyber Security, 2020,5(4): 96-122.)
[73]Abdelghani T. Implementation of defense in depth strategy to secure industrial control system in critical infrastructures[J]. American Journal of Artificial Intelligence, 2019,3(2): 17-22.
[74]游建舟, 呂世超, 孫玉硯, 等. 物聯(lián)網(wǎng)蜜罐綜述[J]. 信息安全學(xué)報, 2020,5(4): 138-156. (You Jianzhou, Lyu Shichao, Sun YuYan, et al. Survey on honeypots of Internet of Things[J]. Journal of Cyber Security, 2020,5(4): 138-156.)
[75]Guan Chongqi, Chen Xianda, Cao Guohong, et al. HoneyCam: scalable high-interaction honeypot for IoT cameras based on 360-degree video[C]//Proc of Conference on Communications and Network Security. Piscataway, NJ: IEEE Press, 2022: 82-90.
[76]Hanson P J, Truax L, Saranchak D D. IoT honeynet for military deception and indications and warnings[C]//Proc of Conference on Autonomous Systems: Sensors, Vehicles, Security, and the Internet of Everything. Bellingham, WA: SPIE Press, 2018: 296-306.
[77]Provos N. A virtual honeypot framework[C]//Proc of the 13th Conference on USENIX Security Symposium. Berkeley, CA: USENIX Association, 2004: 1-14.
[78]Luo Tongbo, Zhao Yanxu, Xin Jin, et al. IoTCandyJar: towards an intelligent-interaction honeypot for IoT devices[J]. Black Hat, 2017, 2017: 1-11.
[79]emic' H, Mrdovic S. IoT honeypot: a multi-component solution for handling manual and Mirai-based attacks[C]//Proc of the 25th Telecommunication Forum. Piscataway, NJ: IEEE Press, 2017: 1-4.
[80]Tabari A Z, Ou Xinming. A multi-phased multi-faceted IoT honeypot ecosystem[C]//Proc of ACM SIGSAC Conference on Computer and Communications Security. New York: ACM Press, 2020: 2121-2123.
[81]Wang Binglai, Dou Yu, Sang Yafei, et al. IoTCMal: towards a hybrid IoT honeypot for capturing and analyzing malware[C]//Proc of IEEE International Conference on Communications. Piscataway, NJ: IEEE Press,2020: 1-7.
[82]Liu Xiaohu, Li Laiqiang, Ma Zhuang, et al. Design of APT attack defense system based on dynamic deception[C]//Proc of the 5th IEEE International Conference on Computer and Communications. Piscataway, NJ: IEEE Press, 2019: 1655-1659.
[83]Guarnizo J D, Tambe A, Bhunia S S, et al. SIPHON: towards scalable high-interaction physical honeypots[C]//Proc of the 3rd ACM Workshop on Cyber-Physical System Security. New York: ACM Press, 2017: 57-68.
[84]王鵑,楊泓遠,樊成陽. 一種基于多階段攻擊響應(yīng)的SDN動態(tài)蜜罐[J]. 信息網(wǎng)絡(luò)安全, 2021, 21(1): 27-40. (Wang Juan, Yang Hongyuan, Fan Chengyang. A SDN dynamic honeypot with multi-phase attack response[J].Netinfo Security, 2021, 21(1): 27-40.)
[85]Hakim M A, Aksu H, Uluagac A S, et al. U-PoT: a honeypot framework for UPnP-based IoT devices[C]//Proc of the 37th IEEE International Performance Computing and Communications Conference. Piscataway, NJ: IEEE Press, 2018: 1-8.
[86]Bai Jun, Xia Jingbo, Zhang Wenjing, et al. Rapid botnet detecting method based on multi-dimensional information divergence[J]. Journal of Huazhong University of Science and Technology, 2014,42(9): 28-32.
[87]原浩宇, 郭軍利, 許明洋. 基于內(nèi)生安全體系結(jié)構(gòu)的蜜罐技術(shù)研究[J]. 計算機應(yīng)用研究, 2023,40(4): 1194-1202. (Yuan Haoyu, Guo Junli, Xu Mingyang. Research on honeypot based on endogenous safety and security architecture[J]. Application Research of Computers, 2023,40(4): 1194-1202.)
[88]Passino F S, Mantziou A, Ghani D, et al. Unsupervised attack pattern detection in honeypot data using Bayesian topic modelling[EB/OL]. (2023-01-06). https://arxiv.org/pdf/2301.02505.pdf.
[89]Valicek M, Schramm G, Pirker M, et al. Creation and integration of remote high interaction honeypots[C]//Proc of International Confe-rence on Software Security and Assurance. Piscataway, NJ: IEEE Press, 2017: 50-55.
[90]Wood T, Ramakrishnan K K, Hwang J, et al. Toward a software-based network: integrating software defined networking and network function virtualization[J]. IEEE Network, 2015, 29(3): 36-41.
[91]Wang He, Wu Bin. SDN-based hybrid honeypot for attack capture[C]//Proc of the 3rd IEEE Information Technology, Networking, Electronic and Automation Control Conference. Piscataway, NJ: IEEE Press, 2019: 1602-1606.
[92]Bubeck S, Chandrasekaran V, Eldan R, et al. Sparks of artificial general intelligence: early experiments with GPT-4[EB/OL]. (2023-04-13). https://arxiv.org/pdf/2303.12712.pdf.
[93]Radford A, Kim J W, Hallacy C, et al. Learning transferable visual models from natural language supervision[C]//Proc of the 38th International Conference on Machine Learning.[S.l.]: PMLR, 2021: 8748-8763.
[94]Ho J, Jain A, Abbeel P. Denoising diffusion probabilistic models[C]//Proc of the 34th International Conference on Neural Information Processing Systems. Red Hook, NY: Curran Associates Inc., 2020: 6840-6851.