Yilong Yang

Membership inference attacks (MIAs) on large language models (LLMs) pose significant privacy risks across various stages of model training. Recent advances in Reinforcement Learning with Verifiable Rewards (RLVR) have brought a profound paradigm shift in LLM training, particularly for complex reasoning tasks. However, the on-policy nature of RLVR introduces a unique privacy leakage pattern: since training relies on self-generated responses without fixed ground-truth outputs, membership inference must now determine whether a given prompt (independent of any specific response) is used during fine-tuning. This creates a threat where leakage arises not from answer memorization. To audit this novel privacy risk, we propose Divergence-in-Behavior Attack (DIBA), the first membership inference framework specifically designed for RLVR. DIBA shifts the focus from memorization to behavioral change, leveraging measurable shifts in model behavior across two axes: advantage-side improvement (e.g., correctness gain) and logit-side divergence (e.g., policy drift). Through comprehensive evaluations, we demonstrate that DIBA significantly outperforms existing baselines, achieving around 0.8 AUC and an order-of-magnitude higher TPR@0.1%FPR. We validate DIBA's superiority across multiple settings--including in-distribution, cross-dataset, cross-algorithm, black-box scenarios, and extensions to vision-language models. Furthermore, our attack remains robust under moderate defensive measures. To the best of our knowledge, this is the first work to systematically analyze privacy vulnerabilities in RLVR, revealing that even in the absence of explicit supervision, training data exposure can be reliably inferred through behavioral traces.

Current adversarial examples (AEs) are typically designed for static models. However, with the wide application of Class-Incremental Learning (CIL), models are no longer static and need to be updated with new data distributed and labeled differently from the old ones. As a result, existing AEs often fail after CIL updates due to significant domain drift. In this paper, we propose SAE to enhance the sustainability of AEs against CIL. The core idea of SAE is to enhance the robustness of AE semantics against domain drift by making them more similar to the target class while distinguishing them from all other classes. Achieving this is challenging, as relying solely on the initial CIL model to optimize AE semantics often leads to overfitting. To resolve the problem, we propose a Semantic Correction Module. This module encourages the AE semantics to be generalized, based on a visual-language model capable of producing universal semantics. Additionally, it incorporates the CIL model to correct the optimization direction of the AE semantics, guiding them closer to the target class. To further reduce fluctuations in AE semantics, we propose a Filtering-and-Augmentation Module, which first identifies non-target examples with target-class semantics in the latent space and then augments them to foster more stable semantics. Comprehensive experiments demonstrate that SAE outperforms baselines by an average of 31.28% when updated with a 9-fold increase in the number of classes.