Russian Developers Teach Neural Networks to “Forget” Faces

Specialists from the IT company Kriptonit have developed a method that removes the digital profile of a specific individual from a neural network’s memory. The technology is linked to the “right to be forgotten” and allows users to withdraw consent for the use of their biometric data. At the same time, the algorithm continues to correctly recognize other users.

An Unreliable “Eraser”

In existing facial-recognition systems, even after a photograph is removed from the database, the neural network can still “remember” the person. The algorithm retains that individual’s digital profile inside the trained model and continues to identify the face, simply not displaying the result. In practice, this means true digital deletion does not occur. This feature creates problems for companies and government systems. Biometric data cannot be changed like a password. If such information leaks, attackers could use it to create deepfakes or gain access to devices, services or financial accounts.

Forgetting Ivan Petrov

Kriptonit’s approach works directly at the model level. The algorithm modifies the internal parameters of the neural network and removes the digital representation of a specific person. After that, the system stops recognizing that individual’s face.

Tests were conducted using open benchmark datasets. The evaluation showed that recognition accuracy for removed users drops to 88%, while the overall performance of the system remains unchanged.

Who Needs to Be Forgotten?

The technology can be used in video-surveillance systems, access-control platforms and corporate security infrastructure. It allows companies to remove biometric profiles of former employees while complying with legal requirements.

Developers are also considering integrating the technology into industrial computer-vision frameworks. This could give companies and government organizations new tools for more responsible handling of biometric data.