Efficient Information Erasure from Artificial Intelligence Systems

Strategic Insights -

The need to remove specific information from artificial intelligence (AI) systems is becoming increasingly important, driven by data privacy concerns, regulatory requirements, and the need to update systems with new data. However, traditional methods of information removal have been challenging due to the complex and intertwined nature of the information within AI systems. Researchers have developed a new approach to efficiently erase information from AI systems, using a combination of retraining and compression techniques.

Key Takeaways:

  • The new method involves retraining AI models using a retraining module to produce an uncorrelated output, followed by compressing the model using tensor networks.
  • The process allows for efficient removal of specific information from AI systems, without significantly affecting their performance.
  • The method can be applied to a variety of AI models, including layered computational models, language processing models, and binary classification models.
  • The retraining module can be used to adapt the model to new data or to remove specific information.
  • The tensorization module uses mathematical structures to compress layers of the computational model, allowing for efficient storage and processing of the compressed model.
  • The method includes a user interface for inputting data to be removed and provides feedback on the progress of retraining and compression.
  • The claims of the patent application cover a range of scenarios, including the use of the method with convolutional operations and attention mechanisms.

Statistics:

  • The patent application was filed on December 28, 2023.
  • The patent application was made available online on June 19, 2025.
  • The application includes 20 claims, covering a range of scenarios for efficient information erasure from AI systems.
  • The method uses a combination of retraining and compression techniques to efficiently remove specific information from AI systems.
  • The retraining module can be used to adapt the model to new data or to remove specific information.
  • The tensorization module uses mathematical structures to compress layers of the computational model, allowing for efficient storage and processing of the compressed model.

Sources:

  • ORUS, Roman. System And Method For Erasing Information From Artificial Intelligence Systems And Related Methods. U.S. Patent Application Number 20250200347, filed December 28, 2023 and posted June 19, 2025.
  • https://ppubs.uspto.gov/pubwebapp/external.html?q=(20250200347)&db=US-PGPUB&type=ids
  • NewsRx LLC. "Efficient Method for Erasing Information from Artificial Intelligence Systems" (2025)