Membership Inference Attacks (MIA) aim to infer whether a target data record has been utilized for model training or not. Prior attempts have quantified the privacy risks of language models (LMs) via MIAs, but there is still no consensus on whether existing MIA algorithms can cause remarkable privacy leakage on practical Large Language Models (LLMs). Existing MIAs designed for LMs can be classified into two categories: reference-free and reference-based attacks. They are both based on the hypothesis that training records consistently strike a higher probability of being sampled. Nevertheless, this hypothesis heavily relies on the overfitting of target models, which will be mitigated by multiple regularization methods and the generalization of LLMs. The reference-based attack seems to achieve promising effectiveness in LLMs, which measures a more reliable membership signal by comparing the probability discrepancy between the target model and the reference model. However, the performance of reference-based attack is highly dependent on a reference dataset that closely resembles the training dataset, which is usually inaccessible in the practical scenario. Overall, existing MIAs are unable to effectively unveil privacy leakage over practical fine-tuned LLMs that are overfitting-free and private. We propose a Membership Inference Attack based on Self-calibrated Probabilistic Variation (SPV-MIA). Specifically, since memorization in LLMs is inevitable during the training process and occurs before overfitting, we introduce a more reliable membership signal, probabilistic variation, which is based on memorization rather than overfitting. Furthermore, we introduce a self-prompt approach, which constructs the dataset to fine-tune the reference model by prompting the target LLM itself. In this manner, the adversary can collect a dataset with a similar distribution from public APIs.

Membership inference attacks (MIAs) aim to infer whether a data point has been used to train a machine learning model. These attacks can be employed to identify potential privacy vulnerabilities and detect unauthorized use of personal data. While MIAs have been traditionally studied for simple classification models, recent advancements in multi-modal pre-training, such as CLIP, have demonstrated remarkable zero-shot performance across a range of computer vision tasks. However, the sheer scale of data and models presents significant computational challenges for performing the attacks. This paper takes a first step towards developing practical MIAs against large-scale multi-modal models. We introduce a simple baseline strategy by thresholding the cosine similarity between text and image features of a target point and propose further enhancing the baseline by aggregating cosine similarity across transformations of the target. We also present a new weakly supervised attack method that leverages ground-truth non-members (e.g., obtained by using the publication date of a target model and the timestamps of the open data) to further enhance the attack. Our evaluation shows that CLIP models are susceptible to our attack strategies, with our simple baseline achieving over $75\%$ membership identification accuracy. Furthermore, our enhanced attacks outperform the baseline across multiple models and datasets, with the weakly supervised attack demonstrating an average-case performance improvement of $17\%$ and being at least $7$X more effective at low false-positive rates. These findings highlight the importance of protecting the privacy of multi-modal foundational models, which were previously assumed to be less susceptible to MIAs due to less overfitting. Our code is available at https://github.com/ruoxi-jia-group/CLIP-MIA.