Training Data Poisoning

Created: 2024-03-12 12:02
#quicknote

One of the Vulnerabilities in LLM-base applications.

Training data poisoning involves the intentional manipulation of data used for either initial model training, fine-tuning, or embedding processes. Attackers aim to introduce vulnerabilities, backdoors, or biases into the model, potentially leading to:

  • Security Compromises: Poisoning can enable unauthorized actions or the leakage of sensitive information.
  • Performance Issues: Deliberate degradation of the model's accuracy or reliability.
  • Ethical Concerns: Introduction of harmful biases that reflect in the model's output.
  • Downstream Risks: Exploitation of vulnerabilities in software dependent on the poisoned model.
  • Reputational Damage: Loss of trust due to security breaches or biased outputs.

Resources

  1. OWASP

Tags

#aisecurity #llm #cybersecurity