As the era of autonomous agents making decisions on behalf of users unfolds, ensuring contextual integrity (CI) - what is the appropriate information to share while carrying out a certain task - becomes a central question to the field. We posit that CI demands a form of reasoning where the agent needs to reason about the context in which it is operating. To test this, we first prompt LLMs to reason explicitly about CI when deciding what information to disclose. We then extend this approach by developing a reinforcement learning (RL) framework that further instills in models the reasoning necessary to achieve CI. Using a synthetic, automatically created, dataset of only 700 examples but with diverse contexts and information disclosure norms, we show that our method substantially reduces inappropriate information disclosure while maintaining task performance across multiple model sizes and families. Importantly, improvements transfer from this synthetic dataset to established CI benchmarks such as PrivacyLens that has human annotations and evaluates privacy leakage of AI assistants in actions and tool calls. Our code is available at: https://github.com/EricGLan/CI-RL

Language models trained on web-scale corpora risk memorizing and exposing sensitive information, prompting the need for effective machine unlearning. Prior methods mainly focus on input queries to suppress sensitive outputs, yet this often fails to eliminate the underlying knowledge and limits scalability. To address this, we propose Corrective Unlearning with Retrieved Exclusions (CURE), a novel unlearning framework that verifies model outputs for leakage and revises them into safe responses. Specifically, CURE employs a lightweight corrector that is applied to the original model to verify whether outputs contain target knowledge and to rewrite them if any leakage is detected. To efficiently handle large-scale unlearning requests, CURE retrieves unlearning targets that are relevant to the initial response and provides them as in-context references to the corrector for detection and conditional revision. By leveraging this retrieval augmentation, the corrector can adapt to new unlearning requests without additional training. Extensive evaluations demonstrate that CURE substantially reduces information leakage, even from indirect queries where prior works fall short, while maintaining response quality and general utility. Moreover, it demonstrates robustness under continual unlearning scenarios, making it practical for real-world applications.

Large language models (LLMs) have advanced natural language processing (NLP) skills such as through next-token prediction and self-attention, but their ability to integrate broad context also makes them prone to incorporating irrelevant information. Prior work has focused on semantic leakage--bias introduced by semantically irrelevant context. In this paper, we introduce expression leakage, a novel phenomenon where LLMs systematically generate sentimentally charged expressions that are semantically unrelated to the input context. To analyse the expression leakage, we collect a benchmark dataset along with a scheme to automatically generate a dataset from free-form text from common-crawl. In addition, we propose an automatic evaluation pipeline that correlates well with human judgment, which accelerates the benchmarking by decoupling from the need of annotation for each analysed model. Our experiments show that, as the model scales in the parameter space, the expression leakage reduces within the same LLM family. On the other hand, we demonstrate that expression leakage mitigation requires specific care during the model building process, and cannot be mitigated by prompting. In addition, our experiments indicate that, when negative sentiment is injected in the prompt, it disrupts the generation process more than the positive sentiment, causing a higher expression leakage rate.

Large language models (LLMs) face significant challenges in ex-ante reasoning, where analysis, inference, or predictions must be made without access to information from future events. Even with explicit prompts enforcing temporal cutoffs, LLMs often generate outputs influenced by internalized knowledge of events beyond the specified cutoff. This paper introduces a novel task and benchmark designed to evaluate the ability of LLMs to reason while adhering to such temporal constraints. The benchmark includes a variety of tasks: stock prediction, Wikipedia event prediction, scientific publication prediction, and Question Answering (QA), designed to assess factual knowledge under temporal cutoff constraints. We use leakage rate to quantify models' reliance on future information beyond cutoff timestamps. Experimental results reveal that LLMs struggle to consistently adhere to temporal cutoffs across common prompting strategies and tasks, demonstrating persistent challenges in ex-ante reasoning. This benchmark provides a potential evaluation framework to advance the development of LLMs' temporal reasoning ability for time-sensitive applications.

Federated Learning (FL) has emerged as a powerful paradigm for collaborative model training while keeping client data decentralized and private. However, it is vulnerable to Data Reconstruction Attacks (DRA) such as "LoKI" and "Robbing the Fed", where malicious models sent from the server to the client can reconstruct sensitive user data. To counter this, we introduce DRArmor, a novel defense mechanism that integrates Explainable AI with targeted detection and mitigation strategies for DRA. Unlike existing defenses that focus on the entire model, DRArmor identifies and addresses the root cause (i.e., malicious layers within the model that send gradients with malicious intent) by analyzing their contribution to the output and detecting inconsistencies in gradient values. Once these malicious layers are identified, DRArmor applies defense techniques such as noise injection, pixelation, and pruning to these layers rather than the whole model, minimizing the attack surface and preserving client data privacy. We evaluate DRArmor's performance against the advanced LoKI attack across diverse datasets, including MNIST, CIFAR-10, CIFAR-100, and ImageNet, in a 200-client FL setup. Our results demonstrate DRArmor's effectiveness in mitigating data leakage, achieving high True Positive and True Negative Rates of 0.910 and 0.890, respectively. Additionally, DRArmor maintains an average accuracy of 87%, effectively protecting client privacy without compromising model performance. Compared to existing defense mechanisms, DRArmor reduces the data leakage rate by 62.5% with datasets containing 500 samples per client.

Large Language Models (LLMs) have demonstrated remarkable capabilities in natural language processing but also pose significant privacy risks by memorizing and leaking Personally Identifiable Information (PII). Existing mitigation strategies, such as differential privacy and neuron-level interventions, often degrade model utility or fail to effectively prevent leakage. To address this challenge, we introduce PrivacyScalpel, a novel privacy-preserving framework that leverages LLM interpretability techniques to identify and mitigate PII leakage while maintaining performance. PrivacyScalpel comprises three key steps: (1) Feature Probing, which identifies layers in the model that encode PII-rich representations, (2) Sparse Autoencoding, where a k-Sparse Autoencoder (k-SAE) disentangles and isolates privacy-sensitive features, and (3) Feature-Level Interventions, which employ targeted ablation and vector steering to suppress PII leakage. Our empirical evaluation on Gemma2-2b and Llama2-7b, fine-tuned on the Enron dataset, shows that PrivacyScalpel significantly reduces email leakage from 5.15\% to as low as 0.0\%, while maintaining over 99.4\% of the original model's utility. Notably, our method outperforms neuron-level interventions in privacy-utility trade-offs, demonstrating that acting on sparse, monosemantic features is more effective than manipulating polysemantic neurons. Beyond improving LLM privacy, our approach offers insights into the mechanisms underlying PII memorization, contributing to the broader field of model interpretability and secure AI deployment.

Federated Learning (FL) emerged as a paradigm for conducting machine learning across broad and decentralized datasets, promising enhanced privacy by obviating the need for direct data sharing. However, recent studies show that attackers can steal private data through model manipulation or gradient analysis. Existing attacks are constrained by low theft quantity or low-resolution data, and they are often detected through anomaly monitoring in gradients or weights. In this paper, we propose a novel data-reconstruction attack leveraging malicious code injection, supported by two key techniques, i.e., distinctive and sparse encoding design and block partitioning. Unlike conventional methods that require detectable changes to the model, our method stealthily embeds a hidden model using parameter sharing to systematically extract sensitive data. The Fibonacci-based index design ensures efficient, structured retrieval of memorized data, while the block partitioning method enhances our method's capability to handle high-resolution images by dividing them into smaller, manageable units. Extensive experiments on 4 datasets confirmed that our method is superior to the five state-of-the-art data-reconstruction attacks under the five respective detection methods. Our method can handle large-scale and high-resolution data without being detected or mitigated by state-of-the-art data reconstruction defense methods. In contrast to baselines, our method can be directly applied to both FedAVG and FedSGD scenarios, underscoring the need for developers to devise new defenses against such vulnerabilities. We will open-source our code upon acceptance.

Large Language Models (LLMs) are trained on massive web-crawled corpora. This poses risks of leakage, including personal information, copyrighted texts, and benchmark datasets. Such leakage leads to undermining human trust in AI due to potential unauthorized generation of content or overestimation of performance. We establish the following three criteria concerning the leakage issues: (1) leakage rate: the proportion of leaked data in training data, (2) output rate: the ease of generating leaked data, and (3) detection rate: the detection performance of leaked versus non-leaked data. Despite the leakage rate being the origin of data leakage issues, it is not understood how it affects the output rate and detection rate. In this paper, we conduct an experimental survey to elucidate the relationship between the leakage rate and both the output rate and detection rate for personal information, copyrighted texts, and benchmark data. Additionally, we propose a self-detection approach that uses few-shot learning in which LLMs detect whether instances are present or absent in their training data, in contrast to previous methods that do not employ explicit learning. To explore the ease of generating leaked information, we create a dataset of prompts designed to elicit personal information, copyrighted text, and benchmarks from LLMs. Our experiments reveal that LLMs produce leaked information in most cases despite less such data in their training set. This indicates even small amounts of leaked data can greatly affect outputs. Our self-detection method showed superior performance compared to existing detection methods.

The growth of low-end hardware has led to a proliferation of machine learning-based services in edge applications. These applications gather contextual information about users and provide some services, such as personalized offers, through a machine learning (ML) model. A growing practice has been to deploy such ML models on the user's device to reduce latency, maintain user privacy, and minimize continuous reliance on a centralized source. However, deploying ML models on the user's edge device can leak proprietary information about the service provider. In this work, we investigate on-device ML models that are used to provide mobile services and demonstrate how simple attacks can leak proprietary information of the service provider. We show that different adversaries can easily exploit such models to maximize their profit and accomplish content theft. Motivated by the need to thwart such attacks, we present an end-to-end framework, SODA, for deploying and serving on edge devices while defending against adversarial usage. Our results demonstrate that SODA can detect adversarial usage with 89% accuracy in less than 50 queries with minimal impact on service performance, latency, and storage.

Federated learning was introduced to enable machine learning over large decentralized datasets while promising privacy by eliminating the need for data sharing. Despite this, prior work has shown that shared gradients often contain private information and attackers can gain knowledge either through malicious modification of the architecture and parameters or by using optimization to approximate user data from the shared gradients. However, prior data reconstruction attacks have been limited in setting and scale, as most works target FedSGD and limit the attack to single-client gradients. Many of these attacks fail in the more practical setting of FedAVG or if updates are aggregated together using secure aggregation. Data reconstruction becomes significantly more difficult, resulting in limited attack scale and/or decreased reconstruction quality. When both FedAVG and secure aggregation are used, there is no current method that is able to attack multiple clients concurrently in a federated learning setting. In this work we introduce LOKI, an attack that overcomes previous limitations and also breaks the anonymity of aggregation as the leaked data is identifiable and directly tied back to the clients they come from. Our design sends clients customized convolutional parameters, and the weight gradients of data points between clients remain separate even through aggregation. With FedAVG and aggregation across 100 clients, prior work can leak less than 1% of images on MNIST, CIFAR-100, and Tiny ImageNet. Using only a single training round, LOKI is able to leak 76-86% of all data samples.