Search for: All records

Machine learning has shown tremendous potential for improving the capabilities of network traffic analysis applications, often outperforming simpler rule-based heuristics. However, ML-based solutions remain difficult to deploy in practice. Many existing approaches only optimize the predictive performance of their models, overlooking the practical challenges of running them against network traffic in real time. This is especially problematic in the domain of traffic analysis, where the efficiency of the serving pipeline is a critical factor in determining the usability of a model. In this work, we introduce CATO, a framework that addresses this problem by jointly optimizing the predictive performance and the associated systems costs of the serving pipeline. CATO leverages recent advances in multi-objective Bayesian optimization to efficiently identify Pareto-optimal configurations, and automatically compiles end-to-end optimized serving pipelines that can be deployed in real networks. Our evaluations show that compared to popular feature optimization techniques, CATO can provide up to 3600× lower inference latency and 3.7× higher zero-loss throughput while simultaneously achieving better model performance. 

Since ZMap’s debut in 2013, networking and security researchers have used the open-source scanner to write hundreds of research papers that study Internet behavior. In addition, ZMap has been adopted by the security industry to build new classes of enterprise security and compliance products. Over the past decade, much of ZMap’s behavior—ranging from its pseudorandom IP generation to its packet construction—has evolved as we have learned more about how to scan the Internet. In this work, we quantify ZMap’s adoption over the ten years since its release, describe its modern behavior (and the measurements that motivated changes), and offer lessons from releasing and maintaining ZMap for future tools. 

In this paper, we introduce Clid, a Transport Layer Security (TLS) client identification tool based on unsupervised learning on domain names from the server name indication (SNI) field. Clid aims to provide some information on a wide range of clients, even though it may not be able to identify a definitive characteristic about each one of the clients. This is a different approach from that of many existing rule-based client identification tools that rely on hardcoded databases to identify granular characteristics of a few clients. Often times, these tools can identify only a small number of clients in a real-world network as their databases grow outdated, which motivates an alternative approach like Clid. For this research, we utilize some 345 million anonymized TLS handshakes collected from a large university campus network. From each handshake, we create a TCP fingerprint – comprising IP flags, time-to-live (TTL), TCP window size, initial sequence number, window size, flags, header length, options, max segment size, and window scaling – that identifies each unique client that corresponds to a physical device on the network. Clid uses Bayesian optimization to find the optimal (in a precise sense that we define later) Density-Based Spatial Clustering of Applications with Noise (DBSCAN) clustering of clients and domain names for a set of TLS connections. Clid maps each client cluster to one or more domain clusters that are most strongly associated with it based on the frequency and exclusivity of their TLS connections. While learning highly associated domain names of a client may not immediately tell us specific characteristics of the client like its the operating system, manufacturer, or TLS configuration, it may serve as a strong first step to doing so. There exists prior work [31, 22] that uses the SNI field for client identification. We evaluate Clid’s performance on various subsets of our captured TLS handshakes and on different parameter settings that affect the granularity of identification results. Our experiments show that Clid is able to identify the single most associated domain cluster (a group of similar domain names in a precise sense that we define in §5.3) for at most 90% of clients in 10,000 TLS connections for a real-world traffic. When one or more domain clusters were allowed to be mapped to a single client cluster, Clid identified such domain names for at least 60% of all clients in all our experiments. 

Virtual Private Networks (VPNs) are increasingly being used to protect online users’ privacy and security. However, there is an ongoing arms race between censors that aim to detect and block VPN usage, and VPN providers that aim to obfuscate their services from these censors. In this paper, we explore the feasibility of a simple, protocol-agnostic VPN detection technique based on identifying encapsulated TCP behaviors in UDP-based tunnels. We derive heuristics to distinguish TCP-over-UDP VPN traffic from plain UDP traffic using RFC-defined TCP behaviors. Our evaluations on realworld traffic show that this technique can achieve a false positive rate (FPR) of 0.11%, an order of magnitude lower than existing machine learning-based VPN detection methods. We suggest defenses to evade our detection technique and encourage VPN providers to proactively defend against such attacks.