EVIL-Encoders

该数据集包含用于为安全漏洞生成 Python 代码的示例。为了使数据集代表真正的漏洞利用，它包括从公共数据库的漏洞利用中提取的代码片段。与以前数据集中的通用 Python 代码不同，实际利用的 Python 代码需要对字节数据进行低级操作以进行混淆（即编码 shellcode）。因此，真正的漏洞利用广泛使用 Python 指令在不同编码器之间转换数据、执行低级算术和逻辑运算以及位级切片，这些在以前的通用 Python 数据集中是找不到的。总的来说，我们构建了一个数据集，其中包含 1,114 个针对漏洞利用的 Python 片段的原始样本及其相应的英语意图。这些示例包括复杂的嵌套指令，这是 Python 编程的典型。为了进行更真实的训练和公平的评估，我们保留了开发人员的原始代码片段并且没有分解它们。我们提供了英语意图来完全描述嵌套指令。为了引导 NMT 模型的训练过程，我们在数据集中包含原始的、面向漏洞利用的片段和来自先前通用 Python 数据集的片段。这使 NMT 模型能够生成可以混合通用指令和面向漏洞利用指令的代码。在 Python 代码生成的几个数据集中，我们选择 Django 数据集，因为它的规模很大。该语料库包含来自 Django Web 应用程序框架的 14,426 对独特的 Python 语句及其对应的英文描述。因此，我们的最终数据集包含 15,540 对独特的 Python 代码片段以及它们在自然语言中的意图。

This dataset contains examples for generating Python code for security vulnerabilities. To better represent real-world exploit scenarios, it includes code snippets extracted from exploits in public vulnerability databases. Unlike the generic Python code found in prior datasets, actual exploit Python code requires low-level operations on byte data for obfuscation (i.e., encoding shellcode). As such, real-world exploits extensively use Python instructions to convert data across different encoders, perform low-level arithmetic and logical operations, and conduct bit-level slicing — functionalities absent from prior generic Python datasets. Overall, we constructed a dataset containing 1,114 raw samples of Python snippets tailored for exploit development, along with their corresponding natural language intent descriptions. These examples feature complex nested instructions, a hallmark of Python programming. To enable more realistic training and fair evaluation, we preserved the original code snippets as written by developers without decomposing them. We provide English intent descriptions to fully characterize these nested instructions. To guide the training process of neural machine translation (NMT) models, we included both the original exploit-oriented snippets and samples from prior generic Python datasets in our corpus. This allows NMT models to generate code that mixes both generic instructions and exploit-specific instructions. Among several datasets for Python code generation, we selected the Django dataset due to its large scale. This corpus contains 14,426 unique pairs of Python statements from the Django web application framework and their corresponding English descriptions. As a result, our final dataset comprises 15,540 unique pairs of Python code snippets and their natural language intent descriptions.

搜集汇总

背景与挑战

背景概述

EVIL-Encoders数据集专注于为安全漏洞生成Python代码，包含1,114个来自真实漏洞利用的代码片段及其英语意图，涵盖低级操作如数据转换和位级切片。为增强训练效果，它还整合了Django数据集的14,426对通用Python语句，总计提供15,540对独特的代码-意图对，用于支持自然语言到代码的翻译模型开发。