Dr. Roland Meier

Recent work has shown that machine learning models can be vulnerable to an adversary crafting targeted inputs designed to cause mispredictions. This is critical in security-related applications such as network intrusion detection systems. While past attacks such as mimicry or gradient-based attacks are able to efficiently generate adversarial examples, they require potentially large input modifications, which is not effective at defeating network flow classifiers. In this work, we show that small modifications to the input (e.g., the traffic that the attacker generates) are enough to manipulate the outcome of a classifier. We focus on minimally evasive adversarial examples to defeat tree-ensemble-based network flow classifiers. We develop an attack that builds on a previous attack introduced by Kantchelian et al. in 2016, which formulates evasion for tree ensembles as a Mixed Integer Linear Program, and which we extend by supporting discrete and categorical features, implementing per-feature evasion costs and modeling interfeature dependencies. This makes our attack more applicable to the network flow classification problem, which typically uses diverse and interdependent input features. We demonstrate our attack on the network flow classifier developed by Känzig et al. in 2019, which was trained to detect command and control (C&C) channels in the Locked Shields cyber defense exercise. Our evaluation shows that minor perturbations of 1 to 4 flow features suffice to successfully fool the classifier. We further retrain the network flow classifier using state-of-the-art adversarial boosting and robust decision tree training published by Chen et al. in 2019. For example, using adversarial boosting, the resulting robust classifier shows a 62.5% increased median evasion distance while achieving equivalent precision and recall on unperturbed samples as the original classifier.

Many public and private organizations use wide-area networks (WANs) to connect their geographically distributed sites. Given that these WANs are often critical for the organization’s operations, their security with respect to confidentiality, integrity, and availability is crucial. A high level of security can be reached if the WAN is built with a dedicated network infrastructure, with the organization operating its own layer-2/3 routing, for example, multiprotocol label switching on top of dedicated fibers or leased lines. Unfortunately, this approach is often slow to deploy, requires high operational effort, and is too expensive for many use cases. A cheaper alternative is to construct the WAN as an overlay network on the infrastructure of public Internet service providers (ISPs), for example, using virtual private network tunnels between the sites. Unfortunately, the security of such a WAN is suboptimal. For instance, traffic analysis attacks (on encrypted traffic) can reveal sensitive information transmitted over these public networks, compromised routers between the sites can alter packets, and network-layer distributed denial-of-service (DDoS) attacks can disrupt connectivity. In this paper, we explore a novel inter-ISP network architecture that provides the desired level of control and security for WAN operators, achieving the best of the two above approaches: strong security properties on a cost-efficient public Internet fabric. Our architecture builds on the SCION next-generation Internet architecture and adds extensions for fine-grained path control, connectivity guarantees in the presence of DDoS attacks, and traffic analysis prevention. With this architecture, WAN operators can build on public layer-3 network connectivity services to deploy secure WANs.

Remote shell sessions via protocols such as SSH are essential for managing systems, deploying applications, and running experiments. However, combined with weak passwords or flaws in the authentication process, remote shell access becomes a major security risk, as it allows an attacker to run arbitrary commands in the name of an impersonated user or even a system administrator. For example, remote shells of weakly protected systems are often exploited in order to build large botnets, to send spam emails, or to launch distributed denial of service attacks. Also, malicious insiders in organizations often use shell sessions to access and transfer restricted data. In this work, we tackle the problem of detecting malicious shell sessions based on session logs, i.e., recorded sequences of commands that were executed over time. Our approach is to classify sessions as benign or malicious by analyzing the sequence of commands that the shell users executed. We model such sequences of commands as n-grams and use them as features to train a supervised machine learning classifier. Our evaluation, based on freely available data and data from our own honeypot infrastructure, shows that the classifier reaches a true positive rate of 99.4% and a true negative rate of 99.7% after observing only four shell commands.

The diversity of applications and devices in enterprise networks combined with large traffic volumes make it inherently challenging to quickly identify malicious traffic. When incidents occur, emergency response teams often lose precious time in reverse-engineering the network topology and configuration before they can focus on malicious activities and digital forensics. In this paper, we present a system that quickly and reliably identifies Command and Control (C&C) channels without prior network knowledge. The key idea is to train a classifier using network traffic from attacks that happened in the past and use it to identify C&C connections in the current traffic of other networks. Specifically, we leverage the fact that – while benign traffic differs – malicious traffic bears similarities across networks (e.g., devices participating in a botnet act in a similar manner irrespective of their location).To ensure performance and scalability, we use a random forest classifier based on a set of computationally-efficient features tailored to the detection of C&C traffic. In order to prevent attackers from outwitting our classifier, we tune the model parameters to maximize robustness. We measure high resilience against possible attacks – e.g.,attempts to camouflaging C&C flows as benign traffic – and packet loss during the inference. We have implemented our approach and we show its practicality on a real use case:Locked Shields, the world’s largest cyber defense exercise. In Locked Shields, defenders have limited resources to protect a large, heterogeneous network against unknown attacks. Using recorded datasets (from 2017 and 2018) from a participating team, we show that our classifier is able to identify C&C channels with 99% precision and over 90% recall in near real time and with realistic resource requirements. If the team had used our system in 2018, it would have discovered 10 out of 12 C&C servers p.p1 in the first hours of the exercise.