// MODEL OPTIMIZATION AND PROMPT SYNTAX TERM

Perturbation

A small, often subtle, change or modification made to data.

TECHNICAL DEFINITION

A small, often imperceptible, modification applied to an input data point, frequently used in the context of adversarial attacks to induce misclassification in AI models or in privacy-preserving techniques like differential privacy to add noise.

BACKGROUND

AI safety is an interdisciplinary field focused on preventing accidents, misuse, or other harmful consequences arising from artificial intelligence systems. It encompasses AI alignment, monitoring AI systems for risks, and enhancing their robustness. The field is particularly concerned with existential risks posed by advanced AI models.

READ MORE ON WIKIPEDIA

SYNONYMS & ALIASES

  • Modification
  • alteration
  • noise
  • distortion

USAGE NOTE

Perturbations are key to creating adversarial examples and implementing differential privacy.

DEVELOPERS

Organizations developing technology related to Perturbation.

  • Google AI / DeepMind

    Conducts extensive research into AI robustness, adversarial machine learning, and understanding how subtle input changes (perturbations) affect large language models and other AI systems, informing prompt engineering best practices.

  • OpenAI

    A leading developer of large language models, OpenAI actively researches and implements techniques to improve model robustness, safety, and alignment, which involves understanding and mitigating the effects of prompt perturbations.

  • Meta AI (FAIR)

    Meta AI conducts foundational research in areas like natural language processing, model robustness, and adversarial AI, which includes studying how small input perturbations can alter model behavior and prompt effectiveness.

  • Microsoft Research

    Engages in research on responsible AI, AI safety, and machine learning interpretability, often exploring how perturbations in prompts or inputs affect the behavior and reliability of AI models.

  • Anthropic

    Focused on AI safety and alignment, Anthropic develops models and methodologies, such as Constitutional AI, that inherently consider the sensitivity of AI responses to various inputs and potential perturbations.

  • Hugging Face

    While primarily a platform for ML models and tools, Hugging Face provides resources and libraries that enable researchers and engineers to experiment with prompt design, fine-tuning, and evaluate model robustness against various input perturbations.

  • IBM Research AI

    IBM Research AI investigates trustworthy AI, including model robustness and explainability, where the analysis of how perturbations impact AI system outputs is a critical area of study for reliable AI engineering and prompt design.

RELATED TERMS IN AI ETHICS & SAFETY