This paper studies extractable memorization: training data that an adversary can efficiently extract by querying a machine learning model without prior knowledge of the training dataset. We show an adversary can extract gigabytes of training data from open-source language models like Pythia or GPT-Neo, semi-open models like LLaMA or Falcon, and closed models like ChatGPT. Existing techniques from the literature suffice to attack unaligned models; in order to attack the aligned ChatGPT, we develop a new divergence attack that causes the model to diverge from its chatbot-style generations and emit training data at a rate 150x higher than when behaving properly. Our methods show practical attacks can recover far more data than previously thought, and reveal that current alignment techniques do not eliminate memorization.

Large language models pretrained on a huge amount of data capture rich knowledge and information in the training data. The ability of data memorization and regurgitation in pretrained language models, revealed in previous studies, brings the risk of data leakage. In order to effectively reduce these risks, we propose a framework DEPN to Detect and Edit Privacy Neurons in pretrained language models, partially inspired by knowledge neurons and model editing. In DEPN, we introduce a novel method, termed as privacy neuron detector, to locate neurons associated with private information, and then edit these detected privacy neurons by setting their activations to zero. Furthermore, we propose a privacy neuron aggregator dememorize private information in a batch processing manner. Experimental results show that our method can significantly and efficiently reduce the exposure of private data leakage without deteriorating the performance of the model. Additionally, we empirically demonstrate the relationship between model memorization and privacy neurons, from multiple perspectives, including model size, training time, prompts, privacy neuron distribution, illustrating the robustness of our approach.

Data privacy has emerged as an important issue as data-driven deep learning has been an essential component of modern machine learning systems. For instance, there could be a potential privacy risk of machine learning systems via the model inversion attack, whose goal is to reconstruct the input data from the latent representation of deep networks. Our work aims at learning a privacy-preserving and task-oriented representation to defend against such model inversion attacks. Specifically, we propose an adversarial reconstruction learning framework that prevents the latent representations decoded into original input data. By simulating the expected behavior of adversary, our framework is realized by minimizing the negative pixel reconstruction loss or the negative feature reconstruction (i.e., perceptual distance) loss. We validate the proposed method on face attribute prediction, showing that our method allows protecting visual privacy with a small decrease in utility performance. In addition, we show the utility-privacy trade-off with different choices of hyperparameter for negative perceptual distance loss at training, allowing service providers to determine the right level of privacy-protection with a certain utility performance. Moreover, we provide an extensive study with different selections of features, tasks, and the data to further analyze their influence on privacy protection.