EVIL-Decoders

这是一个建立在 Shellcode_IA32 之上的汇编数据集，该数据集用于从自然语言描述中自动生成汇编，该数据集包含 3,200 条汇编指令，用英语注释，这些指令是从 IA-32 的 shellcode 中收集并为 Netwide Assembler 编写的（ NASM) 用于 Linux。为了使数据更能代表我们要生成的代码（即完整的漏洞利用，包括要在 shellcode 中交付的解码器），我们用进一步的汇编代码样本丰富了数据集，这些样本取自我们的漏洞利用从公共数据库收集。与之前的数据集不同，新的数据集包含来自实际漏洞利用中使用的真实解码器的汇编代码。最终数据集包含 3,715 对独特的汇编代码片段/英语意图。为了更好地支持开发人员自动生成汇编程序，我们超越了自然语言意图与其相应代码之间的一对一映射。因此，数据集包括 783 行（约占数据集的 21%）的多行代码片段，即生成多行汇编代码的意图，由换行符 (\n) 分隔。这些多行代码段包含许多不同的汇编指令，范围在 2 到 5 之间。This is an assembly dataset built upon Shellcode_IA32, dedicated to automatically generating assembly code from natural language descriptions. It contains 3,200 assembly instructions annotated in English, collected from IA-32 shellcode and written for Netwide Assembler (NASM) on Linux. To better align the dataset with the target code we aim to generate—i.e., complete exploit codes including decoders to be delivered within shellcode—we enriched the dataset with additional assembly code samples sourced from public database-collected exploits. Unlike prior datasets, this new dataset includes assembly code from real decoders used in actual exploit scenarios. The final dataset consists of 3,715 unique pairs of assembly code snippets and English intent descriptions. To better support developers in automatically generating assembly programs, we extended beyond the one-to-one mapping between natural language intents and their corresponding code. As a result, the dataset includes 783 multi-line code snippets, accounting for approximately 21% of the total dataset, which correspond to intents that generate multiple lines of assembly code separated by newline characters ( ). These multi-line snippets contain various assembly instructions, with the number of instructions ranging from 2 to 5 per snippet.

搜集汇总

背景与挑战

背景概述

EVIL-Decoders是一个基于Shellcode_IA32的汇编数据集，用于从英语注释的自然语言描述中自动生成汇编代码。它包含3,715对独特的汇编代码片段与意图，其中21%为多行代码片段，汇编指令来源于IA-32 shellcode和实际漏洞利用中的真实解码器。