Python para hacking de machine learning"- Christian Camilo Urcuqui López

### Overview At Ecoparty, senior data scientist and cybersecurity researcher Cristian Camilo Urcuqui López presented on using Python for adversarial machine learning and machine learning hacking. The presentation emphasizes that machine learning systems are fundamentally statistical models, meaning predictions inevitably contain errors that attackers can systematically exploit. Through open-source Python tools, frameworks, and live demonstrations, the talk outlines how traditional offensive cybersecurity methodologies map directly onto AI and ML pipelines.

### Key Topics - **Intersection of AI and Cybersecurity**: Addressing the complexity in modern security operations (such as SOC alert prioritization) and the transition from reactive, signature-based defense mechanisms to predictive, pattern-based machine learning models. - **Machine Learning Vulnerabilities and Pipeline Attacks**: Identifying attack surfaces across the machine learning lifecycle—including data collection (bias and synthetic data manipulation), pipeline processing attacks, model poisoning, Hugging Face model supply chain injection, and deployment pipeline manipulation. - **The Emerging AI Security Stack**: Examining Microsoft's outlined offensive stack comprising full-stack traditional ethical hacking, adversarial machine learning, and prompt/agent hacking, paired with the necessity of AI ethics committees (addressing fairness, diversity, and explainability). - **Risk Frameworks**: Highlighting evolving risk structures from OWASP (Top 10 for traditional ML and LLMs) and MITRE's tactical components for AI risks.

### Technical Details - **Adaptive Phishing via Reinforcement Learning**: Demonstrating a master's thesis project (co-developed with student Jessica) utilizing Streamlit and a cybersecurity model dubbed "white robot" to dynamically generate phishing and benign samples, using user feedback to refine and personalize attacks through reinforcement learning. - **Nearest Neighbors Feature Manipulation**: Outlining feature-level attacks where an attacker manipulates attributes of a target's nearest neighbors (demonstrated conceptually with a credit approval dataset) to force favorable classification decisions without altering the subject's own profile directly. - **Evasion and Perturbation Attacks on Vision Models**: Introducing the "In Your Face AI" project, which uses deepfake detection models. The speaker demonstrates white-box evasion by multiplying an RGB image matrix by a perturbation mask, altering predictions from fake to 100% real while remaining imperceptible to humans. - **Adversarial Robustness Toolbox (ART)**: Using the ART Python library to execute white-box attacks against neural networks using algorithms such as Carlini & Wagner, SimBA, and HopSkipJump. - **Orion Framework**: Presenting a unified agent-based framework designed around Sun Tzu's *The Art of War* ("Know Your Enemy") that orchestrates attacks across modalities (vision, text, audio) and assists defenders in identifying vulnerabilities and countermeasures.

### Takeaways - Every machine learning model relies on statistical approximations, inherently introducing errors that can be weaponized into adversarial vectors. - Red teams must expand beyond traditional network and software exploitation to include adversarial ML techniques and agent/prompt hacking. - Practitioners are encouraged to explore and translate research papers and mathematical attack formulations into automated open-source Python tooling.