Pingzhi Li

This work proposes ARIA (autoencoder-gated inference-time unlearning), a test-time unlearning method that leaves model weights intact and gates access to unwanted knowledge only when generation enters a forget-related state and introduces three post-unlearning adversarial attacks targeting weight-space and decoding-space recovery.

MLLMs can leak sensitive information from images, exposing new privacy risks that traditional models do not face.

Foundation models can be made intrinsically resistant to unauthorized fine-tuning by concentrating learning in a sparsely masked subnetwork that remains private.