Research Report: ICARUS: Understanding De Facto Formats by Way of Feathers and Wax

Abstract

When a data format achieves a significant level of adoption, the presence of multiple format implementations expands the original specification in often-unforeseen ways. This results in an implicitly defined, de facto format, which can create vulnerabilities in programs handling the associated data files. In this paper we present our initial work on ICARUS: a toolchain for dealing with the problem of understanding and hardening de facto file formats. We show the results of our work in progress in the following areas: labeling and categorizing a corpora of data format samples to understand accepted variations of a format; the detection of sublanguages within the de facto format using both entropy-and taint-tracking-based methods, as a means of breaking down the larger problem of learning how the grammar has evolved; grammar inference via reinforcement learning, as a means of tying together the learned sublanguages; and the defining of both safe subsets of the de facto grammar, as well as translations from unsafe regions of the de facto grammar into safe regions. Real-world data formats evolve as they find use in real-world applications, and a comprehensive ICARUS toolchain for understanding and hardening the resulting de facto formats can identify and address security risks arising from this evolution. 327 2020 Symposium on Security and Privacy Workshops (SPW)

Anomaly Detection with Neural Parsers That Never Reject

“…Given a grammar, we can generate a parser for that grammar, and then apply that parser to a sentence; if the parser fails, then the sentence can be labeled as anomalous. However, it can sometimes be more natural to learn a parser, rather than a grammar, as was done in (Cowger et al, 2020;Woods, 2021). As the resulting parser successfully produces a parse tree for any sentence, it cannot be used directly for anomaly detection.…”

“…For the specific example in Fig. 2 (left), the production rules happen to cover all sentences described by the grammar S -> '{' ('a' | 'b' | 'c' | S+) '}', referred to as the Simple-JSON grammar (Cowger et al, 2020;Woods, 2021); in particular, the '{G}' atom (appearing on the left side of the production rule '{G}' -> '{G' '}') approximately corresponds to S. However, while the rules will successfully parse any valid Simple-JSON sentence, will they also fail at parsing any sentence that is not in the Simple-JSON format, i.e., that is anomalous from the perspective of Simple-JSON? We will address this question in the following two subsections.…”

“…Recently, Reinforcement Learning (RL) has been used as the basis for a novel, flexible grammatical inference scaffolding called RL-GRIT (Cowger et al, 2020;Woods, 2021). The goal of RL is to find a policy, defined as a mapping of observations to actions, which attempts to maximize some reward.…”

“…The primary reason for this is that the learning process benefits from a search space with progressive improvements in the measured reward, rather than the "allor-nothing" design of traditional grammars, which either accept or reject a sentence. While this approach was shown to be promising at automatically generating parsers for formats with data type recurrency (Cowger et al, 2020;Woods, 2021), it also has limitations. First, as is often the case with neural network-based approaches, there is a lack of explainability: the parser has learned some grammar, but it does not provide the production rules for that grammar.…”

RL-GRIT: Reinforcement Learning for Grammar Inference

“…1 were able to be learned automatically, solely from data. Applying RL to the grammar inference problem was initially presented at LangSec 2020 [6],…”