Leaderboards for large reasoning models (LRMs) have turned evaluation into a competition, incentivizing developers to optimize directly on benchmark suites. A shortcut to achieving higher rankings is to incorporate evaluation benchmarks into the training data, thereby yielding inflated performance, known as benchmark contamination. Despite that numerous contamination detection approaches have been proposed, surprisingly, our studies find that evading contamination detections for LRMs is alarmingly easy. We focus on the two scenarios where contamination may occur in practice: (I) when the base model evolves into LRM via supervised fine-tuning (SFT) and reinforcement learning (RL), we find that contamination during SFT can be originally identified by contamination detection methods. Y et, even a brief Group Relative Policy Optimization (GRPO) training can markedly conceal contamination signals that most detection methods rely on. Further empirical experiments and theoretical analysis indicate that Proximal Policy Optimization (PPO) style importance sampling and clipping objectives are the root cause of this detection concealment, indicating that a broad class of RL methods may inherently exhibit similar concealment capability; (II) when SFT contamination with CoT is applied to advanced LRMs as the final stage, most contamination detection methods perform near random guesses. Without exposure to non-members, contaminated LRMs would still have more confidence when responding to those unseen samples that share similar distributions to the training set, and thus, evade existing memorization-based detection methods. Together, our findings reveal the unique vulnerability of LRMs evaluations: Model developers could easily contaminate LRMs to achieve inflated leaderboards performance while leaving minimal traces of contamination, thereby strongly undermining the fairness of evaluation and threatening the integrity of public leaderboards. This underscores the urgent need for advanced contamination detection methods and trustworthy evaluation protocols tailored to LRMs. Our code is available at https://github.com/ASTRAL-Group/ Competition among model developers has intensified as Large Language Models (LLMs) have demonstrated remarkable capabilities in various real-world tasks (Achiam et al., 2023; Wang et al., 2024). The leaderboards for performance are becoming a competitive arena for all state-of-the-art (SOT A) LLMs. However, inadvertently, benchmark samples may appear during LLMs' pre-training due to vast amounts of web-scraped training data. In addition, in the pursuit of publicity, some model developers may even deliberately incorporate benchmark data into their training sets (Sun et al., 2025), resulting in inflated benchmark performance and leaderboard rankings. We refer to this as the benchmark contamination problem in LLMs (Xu et al., 2024; Balloccu et al., 2024).

Membership Inference Attacks (MIAs) have recently been employed to determine whether a specific text was part of the pre-training data of Large Language Models (LLMs). However, existing methods often misinfer non-members as members, leading to a high false positive rate, or depend on additional reference models for probability calibration, which limits their practicality. To overcome these challenges, we introduce a novel framework called Automatic Calibration Membership Inference Attack (ACMIA), which utilizes a tunable temperature to calibrate output probabilities effectively. This approach is inspired by our theoretical insights into maximum likelihood estimation during the pre-training of LLMs. We introduce ACMIA in three configurations designed to accommodate different levels of model access and increase the probability gap between members and non-members, improving the reliability and robustness of membership inference. Extensive experiments on various open-source LLMs demonstrate that our proposed attack is highly effective, robust, and generalizable, surpassing state-of-the-art baselines across three widely used benchmarks. Our code is available at: Github. 1 Introduction Large Language Models (LLMs), pre-trained on massive text corpora, have shown impressive human-level language understanding, reasoning, and decision-making capabilities [4, 28, 1, 23]. However, their tendency to memorize training data also introduces significant ethical and security concerns [14, 31, 2, 21, 22].

Graph neural networks (GNNs) are widely used for graph-structured data but are vulnerable to membership inference attacks (MIAs) in graph classification tasks, which determine if a graph was part of the training dataset, potentially causing data leakage. Existing MIAs rely on prediction probability vectors, but they become ineffective when only prediction labels are available. We propose a Graph-level Label-Only Membership Inference Attack (GLO-MIA), which is based on the intuition that the target model's predictions on training data are more stable than those on testing data. GLO-MIA generates a set of perturbed graphs for target graph by adding perturbations to its effective features and queries the target model with the perturbed graphs to get their prediction labels, which are then used to calculate robustness score of the target graph. Finally, by comparing the robustness score with a predefined threshold, the membership of the target graph can be inferred correctly with high probability. Our evaluation on three datasets and four GNN models shows that GLO-MIA achieves an attack accuracy of up to 0.825, outperforming baseline work by 8.5% and closely matching the performance of probability-based MIAs, even with only prediction labels.

Vision-Language Models (VLMs), built on pre-trained vision encoders and large language models (LLMs), have shown exceptional multi-modal understanding and dialog capabilities, positioning them as catalysts for the next technological revolution. However, while most VLM research focuses on enhancing multi-modal interaction, the risks of data misuse and leakage have been largely unexplored. This prompts the need for a comprehensive investigation of such risks in VLMs. In this paper, we conduct the first analysis of misuse and leakage detection in VLMs through the lens of membership inference attack (MIA). In specific, we focus on the instruction tuning data of VLMs, which is more likely to contain sensitive or unauthorized information. To address the limitation of existing MIA methods, we introduce a novel approach that infers membership based on a set of samples and their sensitivity to temperature, a unique parameter in VLMs. Based on this, we propose four membership inference methods, each tailored to different levels of background knowledge, ultimately arriving at the most challenging scenario. Our comprehensive evaluations show that these methods can accurately determine membership status, e.g., achieving an AUC greater than 0.8 targeting a small set consisting of only 5 samples on LLaVA.

Large Language Models (LLMs) are increasingly used in a variety of applications, but concerns around membership inference have grown in parallel. Previous efforts focus on black-to-grey-box models, thus neglecting the potential benefit from internal LLM information. To address this, we propose the use of Linear Probes (LPs) as a method to detect Membership Inference Attacks (MIAs) by examining internal activations of LLMs. Our approach, dubbed LUMIA, applies LPs layer-by-layer to get fine-grained data on the model inner workings. We test this method across several model architectures, sizes and datasets, including unimodal and multimodal tasks. In unimodal MIA, LUMIA achieves an average gain of 15.71 % in Area Under the Curve (AUC) over previous techniques. Remarkably, LUMIA reaches AUC>60% in 65.33% of cases -- an increment of 46.80% against the state of the art. Furthermore, our approach reveals key insights, such as the model layers where MIAs are most detectable. In multimodal models, LPs indicate that visual inputs can significantly contribute to detect MIAs -- AUC>60% is reached in 85.90% of experiments.

The lack of data transparency in Large Language Models (LLMs) has highlighted the importance of Membership Inference Attack (MIA), which differentiates trained (member) and untrained (non-member) data. Though it shows success in previous studies, recent research reported a near-random performance in different settings, highlighting a significant performance inconsistency. We assume that a single setting doesn't represent the distribution of the vast corpora, causing members and non-members with different distributions to be sampled and causing inconsistency. In this study, instead of a single setting, we statistically revisit MIA methods from various settings with thousands of experiments for each MIA method, along with study in text feature, embedding, threshold decision, and decoding dynamics of members and non-members. We found that (1) MIA performance improves with model size and varies with domains, while most methods do not statistically outperform baselines, (2) Though MIA performance is generally low, a notable amount of differentiable member and non-member outliers exists and vary across MIA methods, (3) Deciding a threshold to separate members and non-members is an overlooked challenge, (4) Text dissimilarity and long text benefit MIA performance, (5) Differentiable or not is reflected in the LLM embedding, (6) Member and non-members show different decoding dynamics.

In the era of large language models (LLMs), detecting pretraining data has been increasingly important due to concerns about fair evaluation and ethical risks. Current methods differentiate members and non-members by designing scoring functions, like Perplexity and Min-k%. In this paper, we first explore the benefits of unseen data, which can be easily collected after the release of the LLM. We find that the perplexities of LLMs perform differently for members and non-members, after fine-tuning with a small amount of previously unseen data. In light of this, we introduce a novel and effective method termed Fine-tuned Score Deviation (FSD), which improves the performance of current scoring functions for pretraining data detection. In particular, we propose to measure the deviation distance of current scores after fine-tuning on a small amount of unseen data within the same domain. In effect, using a few unseen data can largely decrease the scores of all non-members, leading to a larger deviation distance than members. Extensive experiments demonstrate the effectiveness of our method, significantly improving the AUC score on common benchmark datasets across various models. The impressive performance of large language models (LLMs) arises from large-scale pretraining on massive datasets collected from the internet (Achiam et al., 2023; Touvron et al., 2023b). But, model developers are often reluctant to disclose detailed information about the pretraining datasets, raising significant concerns regarding fair evaluation and ethical risks. Specifically, Recent studies reveal that the pretraining corpus may inadvertently include data from evaluation benchmarks (Sainz et al., 2023; Balloccu et al., 2024), making it difficult to assess the practical capability of LLMs. Considering the vast size of the pretraining dataset and the single iteration of pretraining, it has been increasingly important and challenging to detect pretraining data, which determines whether a piece of text is part of the pretraining dataset.

The widespread deployment of large language models (LLMs) has led to impressive advancements, yet information about their training data, a critical factor in their performance, remains undisclosed. Membership inference attacks (MIAs) aim to determine whether a specific instance was part of a target model's training data. However, applying MIAs to LLMs presents unique challenges due to the massive scale of pre-training data and the ambiguous nature of membership. Additionally, creating appropriate benchmarks to evaluate MIA methods is not straightforward, as training and test data distributions are often unknown. In this paper, we introduce EM-MIA, a novel MIA method for LLMs that iteratively refines membership scores and prefix scores via an expectation-maximization algorithm, leveraging the duality that the estimates of these scores can be improved by each other. Membership scores and prefix scores assess how each instance is likely to be a member and discriminative as a prefix, respectively. Our method achieves state-of-the-art results on the WikiMIA dataset. To further evaluate EM-MIA, we present OLMoMIA, a benchmark built from OLMo resources, which allows us to control the difficulty of MIA tasks with varying degrees of overlap between training and test data distributions. We believe that EM-MIA serves as a robust MIA method for LLMs and that OLMoMIA provides a valuable resource for comprehensively evaluating MIA approaches, thereby driving future research in this critical area. Large language models (LLMs) (Brown et al., 2020; Touvron et al., 2023b) have recently emerged as a groundbreaking development and have had a transformative impact in many fields.

Membership inference attacks (MIAs) on diffusion models have emerged as potential evidence of unauthorized data usage in training pre-trained diffusion models. These attacks aim to detect the presence of specific images in training datasets of diffusion models. Our study delves into the evaluation of state-of-the-art MIAs on diffusion models and reveals critical flaws and overly optimistic performance estimates in existing MIA evaluation. We introduce CopyMark, a more realistic MIA benchmark that distinguishes itself through the support for pre-trained diffusion models, unbiased datasets, and fair evaluation pipelines. Through extensive experiments, we demonstrate that the effectiveness of current MIA methods significantly degrades under these more practical conditions. Based on our results, we alert that MIA, in its current state, is not a reliable approach for identifying unauthorized data usage in pre-trained diffusion models. To the best of our knowledge, we are the first to discover the performance overestimation of MIAs on diffusion models and present a unified benchmark for more realistic evaluation. Our code is available on GitHub: \url{https://github.com/caradryanl/CopyMark}.

To assess memorization and information leakage in models, Membership Inference Attacks (MIAs) aim to determine if a data point was part of a model's training set [1]. However, MIAs designed for pre-trained Large Language Models (LLMs) have been largely ineffective [2, 3]. This is primarily because these MIAs, originally developed for classification models, fail to account for the sequential nature of LLMs. Unlike classification models, which produce a single prediction, LLMs generate text token-by-token, adjusting each prediction based on the context of preceding tokens (i.e., prefix). Prior MIAs overlook token-level loss dynamics and the influence of prefixes on next-token predictability, which contributes to memorization.