Privacy Leakage in Language Models

Large language models (LLMs) are increasingly deployed in privacy-critical and personalization-oriented scenarios, yet the role of context length in shaping privacy leakage and personalization effectiveness remains largely unexplored. We introduce a large-scale benchmark, PAPerBench, to systematically study how increasing context length influences both personalization quality and privacy protection in LLMs. The benchmark comprises approximately 29,000 instances with context lengths ranging from 1K to 256K tokens, yielding a total of 377K evaluation questions. It jointly evaluates personalization performance and privacy risks across diverse scenarios, enabling controlled analysis of long-context model behavior. Extensive evaluations across state-of-the-art LLMs reveal consistent performance degradation in both personalization and privacy as context length increases. We further provide a theoretical analysis of attention dilution under context scaling, explaining this behavior as an inherent limitation of soft attention in fixed-capacity Transformers. The empirical and theoretical findings together suggest a general scaling gap in current models -- long context, less focus. We release the benchmark to support reproducible evaluation and future research on scalable privacy and personalization. Code and data are available at https://github.com/SafeRL-Lab/PAPerBench

Large language models (LLMs) can memorize and reproduce training sequences verbatim, raising privacy and copyright concerns. Existing mitigation methods apply interventions uniformly, often degrading performance on the majority of tokens that generalize normally. We study a token-level log-likelihood ratio between a target model and a domain-competent reference with lower exposure to the evaluated sequences. Exposure-controlled experiments support its use for localizing positions associated with sequence-level verbatim recall. Across evaluated settings, a minority of positions carries a disproportionate share of the positive signal, motivating selective activation correction. From the ideal memorization reduction objective, we derive a probe--steer framework which decomposes intervention into a probe that detects memorization-relevant activations and a steer that applies targeted correction only when the probe exceeds a threshold. We propose Gated Subspace Steering (GSS) as its practical instantiation: the optimal probe--steer pair admits a closed-form solution as the leading singular vectors of a gradient-weighted activation matrix. Across four benchmarks, GSS matches or exceeds state-of-the-art memorization reduction while requiring 10--100$\times$ less compute than optimization-based alternatives.

Modern LLMs are increasingly accessed via black-box APIs, requiring users to transmit sensitive prompts, outputs, and fine-tuning data to external providers, creating a critical privacy risk at the API boundary. We introduce AlienLM, a deployable API-only \cradd{exposure-reduction layer that reduces plaintext exposure} by translating text into an Alien Language via a vocabulary-scale bijection, enabling lossless recovery on the client side. Using only standard fine-tuning APIs, Alien Adaptation Training (AAT) adapts target models to operate directly on alienized inputs. Across four LLM backbones and seven benchmarks, AlienLM retains over 81% of plaintext-oracle performance on average, substantially outperforming random-bijection and character-level baselines. Under adversaries with access to model weights, corpus statistics, and learning-based inverse translation, recovery attacks reconstruct fewer than 0.22% of alienized tokens. Our results demonstrate a practical pathway for \cradd{privacy-aware} LLM deployment under API-only access, substantially reducing plaintext exposure while maintaining task performance. Code and data are available at https://github.com/KimJaehee0725/AlienLM.

Agentic Large Language Models (LLMs) are models able to reason, plan, and execute tools over unstructured data. These abilities enable transformative applications in domains spanning personal assistants, financial, and legal domains, etc. While these systems can substantially improve productivity and service quality, effective agency typically requires access to sensitive personal or organizational information. However, this access introduces critical inference-time privacy risks, specifically regarding contextually appropriate information disclosure. While recent studies highlight the inability of agentic LLMs to consistently adhere to privacy norms, existing defenses often rely on auxiliary LLM-based monitors. However, these defenses are expensive and offer limited protection against attacks that are robust to semantic censorship. To contrast this background, this paper proposes a notion of privacy filters based on activation probing. We show that these filters are both computationally efficient and effective for both single-turn and multi-turn conversational settings. Furthermore, this work provides the first systematic investigation into probing model internals across a conversation trajectory, moving beyond static, single-prompt analysis to capture the evolving state of privacy-sensitive interactions, illustrating their applicability across contexts, attack styles, multiple base models, and challenging settings such as mosaic attacks, while providing nuanced insights into the nature of the discovered concepts and modular construction of probes.

Humans are susceptible to undesirable behaviours and privacy leaks under the influence of alcohol. This paper investigates drunk language, i.e., text written under the influence of alcohol, as a driver for safety failures in large language models (LLMs). We investigate three mechanisms for inducing drunk language in LLMs: persona-based prompting, causal fine-tuning, and reinforcement-based post-training. When evaluated on 5 LLMs, we observe a higher susceptibility to jailbreaking on JailbreakBench (even in the presence of defences) and privacy leaks on ConfAIde, where both benchmarks are in English, as compared to the base LLMs as well as previously reported approaches. Via a robust combination of manual evaluation and LLM-based evaluators and analysis of error categories, our findings highlight a correspondence between human-intoxicated behaviour, and anthropomorphism in LLMs induced with drunk language. The simplicity and efficiency of our drunk language inducement approaches position them as potential counters for LLM safety tuning, highlighting significant risks to LLM safety.

Large language models for code (LLM4Code) have greatly improved developer productivity but also raise privacy concerns due to their reliance on open-source repositories containing abundant personally identifiable information (PII). Prior work shows that commercial models can reproduce sensitive PII, yet existing studies largely treat PII as a single category and overlook the heterogeneous risks among different types. We investigate whether distinct PII types vary in their likelihood of being learned and leaked by LLM4Code, and whether this relationship is causal. Our methodology includes building a dataset with diverse PII types, fine-tuning representative models of different scales, computing training dynamics on real PII data, and formulating a structural causal model to estimate the causal effect of learnability on leakage. Results show that leakage risks differ substantially across PII types and correlate with their training dynamics: easy-to-learn instances such as IP addresses exhibit higher leakage, while harder types such as keys and passwords leak less frequently. Ambiguous types show mixed behaviors. This work provides the first causal evidence that leakage risks are type-dependent and offers guidance for developing type-aware and learnability-aware defenses for LLM4Code.

Multimodal large language models (MLLMs) can inadvertently memorize privacy-sensitive information during training. While existing unlearning methods can remove such content, they often severely degrade the model's foundational capabilities, such as general image understanding. This critical shortfall motivates our investigation into benign memory forgetting, the precise removal of targeted, privacy-sensitive knowledge while rigorously preserving unrelated capabilities. To pioneer and evaluate progress toward this objective, we introduce S-MLLMUn Bench, the first benchmark designed to jointly and quantitatively assess an unlearning method's efficacy in knowledge erasure and the preservation of image understanding. Furthermore, we propose the Sculpted Memory Forgetting Adapter (SMFA), a new framework that enables benign memory forgetting. SMFA confines forgetting to designated memory regions, maintaining overall model performance. By initially fine-tuning the model to replace sensitive outputs with refusals, SMFA generates a memory forgetting adapter, followed by a retaining anchor-guided masking mechanism that safeguards unrelated knowledge. Extensive experiments on S-MLLMUn Bench demonstrate that existing methods fail to achieve benign forgetting, whereas our proposed SMFA serves as an effective baseline, successfully achieving targeted knowledge erasure without compromising the model's foundational visual capabilities. Code and data are available at https://github.com/zeng-zhen/S-MLLMUn.

Large language models represent text as hidden numerical patterns, but what information these representations actually contain remains a mystery. Can we recover the original text from just one of them? In this paper, we introduce Rep2Text, a method designed to answer this question by reconstructing text from a single hidden representation. We find that these representations preserve much more information than we might expect. For short texts of about 16 words, Rep2Text can recover roughly half of the original words while preserving their overall meaning. As the text gets longer, recovering the exact wording becomes harder, but the main ideas can still be recovered. These findings offer a new way to understand what language models store internally and highlight potential privacy risks, as hidden representations may reveal parts of the original input.

Large language models (LLMs) are rapidly being adopted for tasks like drafting emails, summarizing meetings, and answering health questions. In these settings, users may need to share private information (e.g., contact details, health records). To evaluate LLMs' ability to identify and redact such information, prior work introduced real-life, scenario-based benchmarks (e.g., ConfAIde, PrivacyLens) and found that LLMs can leak private information in complex scenarios. However, these evaluations relied on proxy LLMs to judge the helpfulness and privacy-preservation quality of LLM responses, rather than directly measuring users' perceptions. To understand how users perceive the helpfulness and privacy-preservation quality of LLM responses to privacy-sensitive scenarios, we conducted a user study ($n=94$) using 90 PrivacyLens scenarios. We found that users had low agreement with each other when evaluating identical LLM responses. In contrast, five proxy LLMs reached high agreement, yet each proxy LLM had low correlation with users' evaluations. These results indicate that proxy LLMs cannot accurately estimate users' wide range of perceptions of utility and privacy in privacy-sensitive scenarios. We discuss the need for more user-centered studies to measure LLMs' ability to help users while preserving privacy, and for improving alignment between LLMs and users in estimating perceived privacy and utility.

Auditing information disclosure across every datapoint during the training of LLMs is challenging. We propose a principled, attack-agnostic approach that uses mutual information to measure what the final model reveals about a datapoint's training membership. We show that this final-model disclosure is upper bounded by the sum of per-iteration gradient disclosures and that, under a reasonable set of assumptions, these gradient disclosures increase with Gradient Uniqueness (GNQ), which measures how distinguishable a datapoint's gradient is relative to other gradients in the batch. While naively computing GNQ requires forming and inverting a $P\times P$ matrix for every datapoint (for a model with $P$ parameters), we introduce Batch-Space Ghost (BS-Ghost). This efficient algorithm performs all computations in a much smaller batch space and uses ghost kernels to compute GNQ "in-run" for every datapoint in the training corpus, with minimal computational and memory overhead. Our experiments show the following: (i) GNQ predicts MIA vulnerability without the need for shadow models. (ii) Beyond membership disclosure, GNQ predicts the success of reconstruction attacks. (iii) GNQ-guided removal and retraining identify datapoints that causally contribute to disclosure. (iv) GNQ outperforms counterfactual memorization in text extraction and common-knowledge discrimination without the need for additional model training. (v) For data attribution, GNQ-guided filtering reduces emergent misalignment in Qwen2.5-7B more than baselines. Further, GNQ attributes 1000 datapoints in 17 seconds---roughly $70\times$ faster than the baselines. (vi) GNQ explains how training choices affect training-set disclosure and captures how per-datapoint disclosure emerges during training.

Recent work shows that distinct inputs produce distinct hidden representations in LLMs, so the input can in principle be recovered from them. Yet, no existing method can reliably do so from the model's output text or logits alone. We introduce Sparse One-hot Discrete Adam (SODA), a white-box search-based algorithm that reconstructs textual inputs from partial knowledge of the output. Given output logits, SODA recovers natural sequences of up to 17 tokens with zero false positives and achieves a 4x improvement in exact-match reconstruction over the strongest prior method. We further show that input length and vocabulary size matter more for inversion difficulty than parameter count, with SODA inverting models from 33M to 3B parameters. Finally, we demonstrate proof-of-concept security applications: SODA recovers 6 backdoor triggers from a poisoned LLM in under 1000 iterations, and 139 tokens of personally identifiable information from synthetic user-chatbot interactions. Our code is available at https://github.com/AdrSkapars/inversion_optimisation.

Retrieval-Augmented Generation (RAG) is a promising technique for applying LLMs to proprietary domains. However, retrieved documents may contain sensitive knowledge, posing risks of privacy leakage in generative results. Thus, effectively erasing private information from retrieved documents is a key challenge for RAG. Unlike traditional text anonymization, RAG should consider: (1) the inherent multi-document reasoning may face de-anonymization attacks; (2) private knowledge varies by scenarios, so users should be allowed to customize which information to erase; (3) preserving sufficient publicly available knowledge for generation tasks. This paper introduces the privacy erasure task for RAG and proposes Eraser4RAG, a private knowledge eraser which effectively removes user-defined private knowledge from documents while preserving sufficient public knowledge for generation. Specifically, we first construct a global knowledge graph to identify potential knowledge across documents, aiming to defend against de-anonymization attacks. Then we randomly split it into private and public sub-graphs, and fine-tune Flan-T5 to rewrite the retrieved documents excluding private triples. Finally, PPO algorithm optimizes the rewriting model to minimize private triples and maximize public triples retention. Experiments on four QA datasets demonstrate that Eraser4RAG achieves superior erase performance than GPT-4o.

Language models (LMs) can "memorize" information, i.e., encode training data in their weights in such a way that inference-time queries can lead to verbatim regurgitation of that data. This ability to extract training data can be problematic, for example, when data are private or sensitive. In this work, we investigate methods to mitigate memorization: three regularizer-based, three finetuning-based, and eleven machine unlearning-based methods, with five of the latter being new methods that we introduce. We also introduce TinyMem, a suite of small, computationally-efficient LMs for the rapid development and evaluation of memorization-mitigation methods. We demonstrate that the mitigation methods that we develop using TinyMem can successfully be applied to production-grade LMs, and we determine via experiment that: regularizer-based mitigation methods are slow and ineffective at curbing memorization; fine-tuning-based methods are effective at curbing memorization, but overly expensive, especially for retaining higher accuracies; and unlearning-based methods are faster and more effective, allowing for the precise localization and removal of memorized information from LM weights prior to inference. We show, in particular, that our proposed unlearning method BalancedSubnet outperforms other mitigation methods at removing memorized information while preserving performance on target tasks.

Instruction tuning aligns large language models (LLMs) with human intentions but requires diverse, high-quality data that are difficult to collect in privacy-sensitive domains. Federated instruction tuning (FedIT) enables collaborative training across data owners, yet existing methods typically assume sufficient local data. In realistic few-shot settings, limited samples can cause overfitting, degrade performance, and increase vulnerability to training data extraction attacks. We propose PPFedIT, a federated algorithm that improves both model performance and privacy protection in federated few-shot learning. It comprises three client-side steps: (1) synthetic data generation, which uses LLMs to diversify and enrich local data; (2) parameter isolation training, which updates the shared global LLM on synthetic data and local LLMs on private local data to mitigate synthetic-data noise; and (3) local aggregation then sharing, which mixes global and local model parameters before uploading them for server aggregation to mitigate data extraction attacks. Experiments on three open-source datasets show that PPFedIT improves model performance by an average of 8.4% and reduces the risk of data extraction attacks by approximately 20% in challenging federated few-shot settings.