视频流量逐渐在网络中占据主导地位,且视频平台大多对其进行加密传输。虽然加密传输视频可以有效保护用户隐私,但是也增加了监管有害视频传播的难度.现有的加密视频识别方法基于TCP(Transmission Control Protocol)传输协议头部信息和HT...视频流量逐渐在网络中占据主导地位,且视频平台大多对其进行加密传输。虽然加密传输视频可以有效保护用户隐私,但是也增加了监管有害视频传播的难度.现有的加密视频识别方法基于TCP(Transmission Control Protocol)传输协议头部信息和HTTP/1.1(Hypertext Transfer Protocol Version1.1)的传输模式,提取应用层音视频数据单元传输长度序列来实现视频识别.但是随着基于UDP(User Datagram Protocol)的QUIC(Quick UDP Internet Connections)协议及基于QUIC实现的HTTP/3(Hypertext Transfer Protocol Version 3)协议应用于视频传输,已有方法不再适用.HTTP/3协议缺少类似TCP的头部信息,且使用了多路复用机制,并对几乎所有数据进行了加密,此外,视频平台开始使用多片段合并分发技术,这给从网络流量中精准识别加密视频带来了巨大挑战。本文基于HTTP/3协议中的控制信息特征,提出了从HTTP/3加密视频流中提取数据传输特征并进行修正的方法,最大程度复原出应用层音视频长度特征.面向多片段合并分发导致的海量匹配问题,本文基于明文指纹库设计了键值数据库来实现视频的快速识别.实验结果表明,本文提出的基于HTTP/3传输特性的加密视频识别方法能够在包含36万个真实视频指纹的YouTube大规模指纹库中达到接近99%的准确率,100%的精确率以及99.32%的F1得分,对传输过程中加人了填充顿的Facebook平台,在包含28万个真实视频指纹的大规模指纹库中达到95%的准确率、100%的精确率以及96.45%的F1得分,在具有同样特性的Instagram平台中,最高可达到97.57%的F1得分,且本方法在所有指纹库中的平均视频识别时间均低于0.4秒.本文的方法首次解决了使用HTTP/3传输的加密视频在大规模指纹库场景中的识别问题,具有很强的实用性和通用性.展开更多
In recent years,network attacks have been characterized by diversification and scale,which indicates a requirement for defense strategies to sacrifice generalizability for higher security.As the latest theoretical ach...In recent years,network attacks have been characterized by diversification and scale,which indicates a requirement for defense strategies to sacrifice generalizability for higher security.As the latest theoretical achievement in active defense,mimic defense demonstrates high robustness against complex attacks.This study proposes a Function-aware,Bayesian adjudication,and Adaptive updating Mimic Defense(FBAMD)theory for addressing the current problems of existing work including limited ability to resist unknown threats,imprecise heterogeneous metrics,and over-reliance on relatively-correct axiom.FBAMD incorporates three critical steps.Firstly,the common features of executors’vulnerabilities are obtained from the perspective of the functional implementation(i.e,input-output relationships extraction).Secondly,a new adjudication mechanism considering Bayes’theory is proposed by leveraging the advantages of both current results and historical confidence.Furthermore,posterior confidence can be updated regularly with prior adjudication information,which provides mimic system adaptability.The experimental analysis shows that FBAMD exhibits the best performance in the face of different types of attacks compared to the state-of-the-art over real-world datasets.This study presents a promising step toward the theo-retical innovation of mimic defense.展开更多
文摘视频流量逐渐在网络中占据主导地位,且视频平台大多对其进行加密传输。虽然加密传输视频可以有效保护用户隐私,但是也增加了监管有害视频传播的难度.现有的加密视频识别方法基于TCP(Transmission Control Protocol)传输协议头部信息和HTTP/1.1(Hypertext Transfer Protocol Version1.1)的传输模式,提取应用层音视频数据单元传输长度序列来实现视频识别.但是随着基于UDP(User Datagram Protocol)的QUIC(Quick UDP Internet Connections)协议及基于QUIC实现的HTTP/3(Hypertext Transfer Protocol Version 3)协议应用于视频传输,已有方法不再适用.HTTP/3协议缺少类似TCP的头部信息,且使用了多路复用机制,并对几乎所有数据进行了加密,此外,视频平台开始使用多片段合并分发技术,这给从网络流量中精准识别加密视频带来了巨大挑战。本文基于HTTP/3协议中的控制信息特征,提出了从HTTP/3加密视频流中提取数据传输特征并进行修正的方法,最大程度复原出应用层音视频长度特征.面向多片段合并分发导致的海量匹配问题,本文基于明文指纹库设计了键值数据库来实现视频的快速识别.实验结果表明,本文提出的基于HTTP/3传输特性的加密视频识别方法能够在包含36万个真实视频指纹的YouTube大规模指纹库中达到接近99%的准确率,100%的精确率以及99.32%的F1得分,对传输过程中加人了填充顿的Facebook平台,在包含28万个真实视频指纹的大规模指纹库中达到95%的准确率、100%的精确率以及96.45%的F1得分,在具有同样特性的Instagram平台中,最高可达到97.57%的F1得分,且本方法在所有指纹库中的平均视频识别时间均低于0.4秒.本文的方法首次解决了使用HTTP/3传输的加密视频在大规模指纹库场景中的识别问题,具有很强的实用性和通用性.
基金supported by the National Key Research and Development Program of China(Grant No.2020YFB1804604).
文摘In recent years,network attacks have been characterized by diversification and scale,which indicates a requirement for defense strategies to sacrifice generalizability for higher security.As the latest theoretical achievement in active defense,mimic defense demonstrates high robustness against complex attacks.This study proposes a Function-aware,Bayesian adjudication,and Adaptive updating Mimic Defense(FBAMD)theory for addressing the current problems of existing work including limited ability to resist unknown threats,imprecise heterogeneous metrics,and over-reliance on relatively-correct axiom.FBAMD incorporates three critical steps.Firstly,the common features of executors’vulnerabilities are obtained from the perspective of the functional implementation(i.e,input-output relationships extraction).Secondly,a new adjudication mechanism considering Bayes’theory is proposed by leveraging the advantages of both current results and historical confidence.Furthermore,posterior confidence can be updated regularly with prior adjudication information,which provides mimic system adaptability.The experimental analysis shows that FBAMD exhibits the best performance in the face of different types of attacks compared to the state-of-the-art over real-world datasets.This study presents a promising step toward the theo-retical innovation of mimic defense.
