g5-eval-dashboard / PROJECT_CONTEXT.md
juandavidrodriguezar's picture
Upload folder using huggingface_hub
eaae571 verified
|
Raw
History Blame
2.91 kB

LLM Prompt Injection Security Evaluation Framework

Project Overview

Project Type

  • University research capstone project focused on LLM security.

Project Objective

  • Design and implement a testing framework that evaluates how vulnerable LLM chatbots are to prompt injection attacks.
  • Evaluate how mitigation strategies improve security outcomes.

The system will:

  1. Simulate prompt injection attacks.
  2. Measure how often attacks succeed.
  3. Apply mitigation strategies.
  4. Measure how much mitigation reduces successful attacks.

Scope Clarification

This Project Is NOT

  • Training LLM models
  • Building a new AI model
  • Building a ChatGPT-like system
  • Performing large-scale machine learning training

This Project IS

  • Designing attack scenarios
  • Evaluating LLM security vulnerabilities
  • Implementing defensive mitigation strategies
  • Measuring improvements after mitigation

Attack Scenarios to Evaluate

  • Direct prompt injection through user inputs
  • Attempts to extract hidden system prompts or internal instructions
  • Indirect prompt injection through retrieved documents (RAG scenarios)
  • Adaptive attackers who modify prompts after observing system responses

The framework should allow easily adding new attack strategies.

Mitigation Strategies

  • Prompt hardening
  • Input filtering
  • Instruction isolation
  • Detection mechanisms (e.g., classifier-based detection)

The framework should support comparing multiple mitigation approaches.

The system should also allow analyzing trade-offs such as:

  • Security improvement
  • Usability impact
  • False positives or false negatives

Evaluation Metrics

The system should measure:

  • Attack Success Rate (ASR)
  • Reduction in successful attacks after mitigation
  • False positive / false negative rates for detection methods
  • Impact on normal system performance

Models

The framework may test different LLMs including:

  • API-based models (e.g., GPT-family models)
  • Open-source models (e.g., LLaMA, Mistral)

The architecture should support easily adding new models.

System Architecture (High Level)

The pipeline will conceptually look like this:

LLM Chatbot

Attack Generator

Attack Execution

Mitigation Layer

Evaluation Metrics

Results & Analysis

The framework should support running multiple experiments and logging results.

Repository Goals

The repository should prioritize:

  • Modular architecture
  • Reproducible experiments
  • Clear documentation
  • Experiment logging
  • Easy extension for new attacks
  • Easy extension for new mitigation strategies

Research Inspiration

This project is inspired by research on prompt injection attacks and LLM security, including academic papers that:

  • Formalize prompt injection attacks
  • Benchmark attack strategies
  • Compare defense mechanisms

The goal is to build a practical experimental framework inspired by this research.