HackThisAI - 对抗性机器学习（ AML ）捕获FLA

项目说明

HackThisAI

!logo

This repository is no longer maintained. The challenges may still be useful for some folks, but the concept has e. The format went on to shape two DEFCON/Kaggle CTF competitions and inspired a much higher-quality offering. Contact for more information.

HackThisAI is a series of capture the flag (CTF) challenges to educate and train ad machine learning techniques. CTFs are a common way information security professionals learn new skills and compete across the community. Challenges are categorized so that players can focus on specific topics. The HackThisAI challenges should be approachable for new data scientists, machine learning engineers, and traditional infosec professionals, but progress in difficulty to a point that requires combinations of ad techniques and creati.

Categories

Often, you'll see ad machine learning tasks classified into these buckets. It's a useful framework for bringing an ad mindset to AI/ML products, but remember that these tasks are not mutually exclusi objectives may require combinations of methods that traditionally belong in these categories.

1. Evade the model. Can you evade or fool the model? In many applications, model classification is used in control flow: if model.predict() == A: do this; else: do that. If you can successfully control the model output, you may have some sway o flow. E is also a fundamental task that may help with influence and steal challenges.
2. Influence the model. Can you change how the model operates? You may ha or indirect access to training data or model training steps.
3. Steal the model. Can you steal, extract, or invert the model? In traditional hacking, you might want to use access to exfiltrate a model binary. However, in these challenges, you'll want to interact with the model API to learn things about it's structure and find a way to recreate it.
4. Infer membership. Was this data point used in the model training process? If you can determine that, you could privacy considerations or start learning things about the training dataset.

Playing with the challenges

To lower the barrier to entry, I' most of these challenges using jupyterlite. This is still very experimental, but loads e in your browser with Web Asssembly. The benefit is that you don't need to bother with downloading and en. The downside is that it may be slow/buggy. If your kernel doesn't seem to be working, give it some time. If the jupyterlite interface gets too annoying, you can download the repository or notebooks and run them in your own JupyterLab deployment. The requirements are in en.yml.

Each challenge is self contained in a notebook. Na to the challenge folder and open the notebook with the same name to begin.

I've only ported a portion of the challenges to JupyterLite so far (the ones with fewer/simpler dependencies). Check out GitHub for others.

bad2good

baseball

credit_check

learning_to_fly

Updates

Follow to hear about new challenges, substanti, or any future li CTF events. File any issues on GitHub.

项目信息

开发语言：Jupyter Notebook

Star 数：119

技术要点

该项目属于CTF思路方向的开源资源，代码结构清晰，适合安全研究人员学习参考。项目围绕Adversarial Machine Learning (AML) Capture the Flag (CTF)展开，提供了可运行的实现，对理解相关安全机制、工具原理与攻防技术具有较高的研究价值。

从技术实现角度看，此类项目的核心在于对底层原理的深入理解与工程化落地。开发者在研究过程中可以结合官方文档与社区讨论，梳理项目的初始化流程、核心模块职责与数据流转，逐步掌握其中的技术细节与设计思路。

使用与学习建议

1. 下载源码压缩包，解压到本地环境。

2. 根据项目类型准备对应的运行环境与依赖。

3. 按照 README 中的步骤完成配置与部署。

4. 在隔离的测试环境中运行验证，避免影响生产系统。

免责声明

本资源来源于开源社区，仅供安全研究、技术学习与合法授权范围内的攻防演练使用，请勿用于任何非法或侵害他人权益的用途。使用本项目产生的一切后果由使用者自行承担。