System auditing provides a low-level view into cyber threats by monitoring system entity interactions. In response to advanced cyber-attacks, one prevalent solution is to apply data provenance analysis on audit records to search for anomalies (anomalous behaviors) or specifications of known attacks. However, existing approaches suffer from several limitations: 1) generating high volumes of false alarms, 2) relying on expert knowledge, or 3) producing coarse-grained detection signals. In this paper, we recognize the structural similarity between threat detection in cybersecurity and recommendation in information retrieval. By mapping security concepts of system entity interactions to recommendation concepts of user-item interactions, we identify cyber threats by predicting the preferences of a system entity on its interactive entities. Furthermore, inspired by the recent advances in modeling high-order connectivity via item side information in the recommendation, we transfer the insight to cyber threat analysis and customize an automated detection system, SHADEWATCHER. It fulfills the potential of high-order information in audit records via graph neural networks to improve detection effectiveness. Besides, we equip SHADEWATCHER with dynamic updates towards better generalization to false alarms. In our evaluation against both real-life and simulated cyber-attack scenarios, SHADEWATCHER shows its advantage in identifying threats with high precision and recall rates. Moreover, SHADEWATCHER is capable of pinpointing threats from nearly a million system entity interactions within seconds.

Despite the functional success of deep Neural Networks, their trustworthiness remains a crucial open challenge. To address this challenge, both testing and verification techniques have been proposed. But these existing techniques provide either scalability to large networks or formal guarantees, not both. In this paper, we propose a scalable quantitative verification framework for deep Neural Networks, i.e., a test-driven approach that comes with formal guarantees that a desired probabilistic property is satisfied. Our technique performs enough tests until soundness of a formal probabilistic property can be proven. It can be used to certify properties of both deterministic and randomized DNNs. We implement our approach in a tool called PROVERO and apply it in the context of certifying adversarial robustness of DNNs. In this context, we first show a new attack-agnostic measure of robustness which offers an alternative to purely attack-based methodology of evaluating robustness being reported today. Second, PROVERO provides certificates of robustness for large DNNs, where existing state-of-the-art verification tools fail to produce conclusive results. Our work paves the way forward for verifying properties of distributions captured by real-world deep Neural Networks, with provable guarantees, even where testers only have black-box access to the neural network.

Fuzzing is an effective method to identify bugs and security vulnerabilities in software. One particular difficulty faced by fuzzing is how to effectively generate inputs to cover program paths, especially for programs with complex logic. We observe that complex programs are often composed of components, which is a natural result of software engineering principles. The components interface with each other using memory buffers, forming stages of processing in the program logic. Program logic in later stages is difficult to reach by fuzzers. In this paper, we develop a novel solution to fuzz such program logic, called STAGEFUZZER. It identifies the stages and memory interfaces from program binaries, and fuzzes later stages of the program effectively. In our evaluation with a suite of typical binaries, STAGEFUZZER correctly identifies the program structure and effectively increases the coverage of program logic compared to AFL fuzzer.

Process-based isolation, suggested by several research prototypes, is a cornerstone of modern browser security architectures. Google Chrome is the first commercial browser that adopts this architecture. Unlike several research prototypes, Chrome's process-based design does not isolate different web origins, but primarily promises to protect "the local system" from "the web". However, as billions of users now use web-based cloud services (e.g., Dropbox and Google Drive), which are integrated into the local system, the premise that browsers can effectively isolate the web from the local system has become questionable. In this paper, we argue that, if the process-based isolation disregards the same-origin policy as one of its goals, then its promise of maintaining the "web/local system (local)" separation is doubtful. Specifically, we show that existing memory vulnerabilities in Chrome's renderer can be used as a stepping-stone to drop executables/scripts in the local file system, install unwanted applications and misuse system sensors. These attacks are purely data-oriented and do not alter any control flow or import foreign code. Thus, such attacks bypass binary-level protection mechanisms, including ASLR and in-memory partitioning. Finally, we discuss various full defenses and present a possible way to mitigate the attacks presented.

New hardware primitives such as Intel SGX secure a user-level process in presence of an untrusted or compromised OS. Such "enclaved execution" systems are vulnerable to several side-channels, one of which is the page fault channel. In this paper, we show that the page fault side-channel has sufficient channel capacity to extract bits of encryption keys from commodity implementations of cryptographic routines in OpenSSL and Libgcrypt -- leaking 27% on average and up to 100% of the secret bits in many case-studies. To mitigate this, we propose a software-only defense that masks page fault patterns by determinising the program's memory access behavior. We show that such a technique can be built into a compiler, and implement it for a subset of C which is sufficient to handle the cryptographic routines we study. This defense when implemented generically can have significant overhead of up to 4000X, but with help of developer-assisted compiler optimizations, the overhead reduces to at most 29.22% in our case studies. Finally, we discuss scope for hardware-assisted defenses, and show one solution that can reduce overheads to 6.77% with support from hardware changes.

As control-flow hijacking defenses gain adoption, it is important to understand the remaining capabilities of adversaries via memory exploits. Non-control data exploits are used to mount information leakage attacks or privilege escalation attacks program memory. Compared to control-flow hijacking attacks, such non-control data exploits have limited expressiveness, however, the question is: what is the real expressive power of non-control data attacks? In this paper we show that such attacks are Turing-complete. We present a systematic technique called data-oriented programming (DOP) to construct expressive non-control data exploits for arbitrary x86 programs. In the experimental evaluation using 9 programs, we identified 7518 data-oriented x86 gadgets and 5052 gadget dispatchers, which are the building blocks for DOP. 8 out of 9 real-world programs have gadgets to simulate arbitrary computations and 2 of them are confirmed to be able to build Turing-complete attacks. We build 3 end-to-end attacks to bypass randomization defenses without leaking addresses, to run a network bot which takes commands from the attacker, and to alter the memory permissions. All the attacks work in the presence of ASLR and DEP, demonstrating how the expressiveness offered by DOP significantly empowers the attacker.