// MODEL OPTIMIZATION AND PROMPT SYNTAX TERM

Model Stealing

An attacker tries to copy or replicate a proprietary AI model by observing its behavior, similar to model extraction.

TECHNICAL DEFINITION

A broad category of attacks where an adversary aims to illegally obtain a functional equivalent or the underlying parameters of a proprietary AI model, often through querying its API and inferring its logic, infringing on intellectual property.

BACKGROUND

Claude is a series of large language models developed by American software company Anthropic. Claude was released as an AI-based chatbot in March 2023. It is also used in AI-assisted software development.

READ MORE ON WIKIPEDIA

SYNONYMS & ALIASES

  • Model extraction
  • intellectual property theft
  • model replication

USAGE NOTE

Model stealing is a significant concern for companies that invest heavily in developing advanced AI models.

DEVELOPERS

Organizations developing technology related to Model Stealing.

  • Robust Intelligence

    Develops an 'AI Firewall' that protects machine learning models in production from various threats, including model stealing and data extraction attacks, by validating inputs and outputs in real-time.

  • HiddenLayer

    An AI security company providing a platform that detects and responds to adversarial attacks against machine learning models. Their solutions are designed to protect against threats like model stealing and adversarial examples.

  • Trail of Bits

    A cybersecurity research and consulting firm that offers AI security assurance services. They perform security reviews and adversarial testing on AI systems to identify vulnerabilities that could lead to model stealing.

  • Scale AI

    While primarily known for data annotation, Scale AI offers comprehensive Test & Evaluation (T&E) services for language models. This includes red teaming to assess vulnerabilities such as susceptibility to model extraction attacks.

  • Google AI

    As a major developer of large-scale models, Google's research teams are actively working on AI safety and security. They publish research and develop techniques, such as differential privacy and watermarking, to defend against model stealing attacks.

  • Microsoft Research

    Microsoft's research division focuses on responsible AI, which includes security and privacy. They develop frameworks and tools to build secure AI systems and research countermeasures for model extraction and other adversarial attacks.

  • MIT Lincoln Laboratory

    A federally funded research center that works on technology for national security. Their AI research includes developing robust and secure machine learning systems, which involves studying and creating defenses against model extraction threats.

  • Preamble

    Provides an AI security platform to evaluate, monitor, and secure large language models against risks. Their tools test for vulnerabilities including data leakage and model extraction to ensure safe deployment.

RELATED TERMS IN AI ETHICS & SAFETY