Stateful firewalls (SFW) play a critical role in securing our network infrastructure. Incorrect implementation of the intended stateful semantics can lead to evasion opportunities, even if firewall rules are configured correctly. Uncovering these opportunities is challenging due to the (1) black-box and proprietary nature of firewalls; (2) diversity of deployments; and (3) complex stateful semantics. To tackle these challenges, we present Pryde. Pryde uses a modular model-guided workflow that generalizes across black-box firewall implementations and deployment-specific settings to generate evasion attacks. Pryde infers a behavioral model of the stateful firewall in the presence of potentially non-TCP-compliant packet sequences. It uses this model in conjunction with attacker capabilities and victim behavior to synthesize custom evasion attacks. Using Pryde, we identify more than 6,000 unique attacks against 4 popular firewalls and 4 host networking stacks, many of which cannot be uncovered by prior work on censorship circumvention and black-box fuzzing.

Modern systems across diverse application domains (e.g., IoT, automotive) have many black-box devices whose internal structures and/or protocol formats are unknown. We currently lack the tools to systematically understand the behavior and learn the security weaknesses of these black-box devices. Such tools could enable many use cases, such as: 1) identifying input packets that lead to network attacks; and 2) inferring the format of unknown protocols. Our goal is to enable oblivious network analysis which can perform the aforementioned tasks for black-box devices. In this work, we explore the use of a recent machine learning tool called generative adversarial networks (GANs) [16] to enable this vision. Unlike other competing approaches, GANs can work in a truly black-box setting and can infer complex dependencies between protocol fields with little to no supervision. We leverage GANs to show the preliminary use cases of our approaches using two case studies: 1) generating synthetic protocol messages given only samples of messages; and 2) generating attack inputs for a black-box system. While there are still many open challenges, our results suggest the early promise of GANs to enable "oblivious" analysis of networked elements.

Recent studies have shown a range of co-residency side channels that can be used to extract private information from cloud clients. Unfortunately, addressing these side channels often requires detailed attack-specific fixes that require significant modifications to hardware, client virtual machines (VM), or hypervisors. Furthermore, these solutions cannot be generalized to future side channels. Barring extreme solutions such as single tenancy which sacrifices the multiplexing benefits of cloud computing, such side channels will continue to affect critical services. In this work, we present Nomad, a system that offers vector-agnostic defense against known and future side channels. Nomad envisions a provider-assisted VM migration service, applying the moving target defense philosophy to bound the information leakage due to side channels. In designing Nomad, we make four key contributions: (1) a formal model to capture information leakage via side channels in shared cloud deployments; (2) identifying provider-assisted VM migration as a robust defense for arbitrary side channels; (3) a scalable online VM migration heuristic that can handle large datacenter workloads; and (4) a practical implementation in OpenStack. We show that Nomad is scalable to large cloud deployments, achieves near-optimal information leakage subject to constraints on migration overhead, and imposes minimal performance degradation for typical cloud applications such as web services and Hadoop MapReduce.