The substantial investment required to develop Large Language Models (LLMs) makes them valuable intellectual property, raising significant concerns about copyright protection. LLM fingerprinting has emerged as a key technique to address this, which aims to verify a model's origin by extracting an intrinsic, unique signature (a ''fingerprint'') and comparing it to that of a source model to identify illicit copies. However, existing black-box fingerprinting methods often fail to generate distinctive LLM fingerprints. This ineffectiveness arises because black-box methods typically rely on model outputs, which lose critical information about the model's unique parameters due to the usage of non-linear functions. To address this, we first leverage Fisher Information Theory to formally demonstrate that the gradient of the model's input is a more informative feature for fingerprinting than the output. Based on this insight, we propose ZeroPrint, a novel method that approximates these information-rich gradients in a black-box setting using zeroth-order estimation. ZeroPrint overcomes the challenge of applying this to discrete text by simulating input perturbations via semantic-preserving word substitutions. This operation allows ZeroPrint to estimate the model's Jacobian matrix as a unique fingerprint. Experiments on the standard benchmark show ZeroPrint achieves a state-of-the-art effectiveness and robustness, significantly outperforming existing black-box methods.

While retrieval-augmented generation systems partially address the hallucination issues in large language models, it also introduces new vulnerabilities to knowledge corruption attacks. Adversaries exploit these vulnerabilities by poisoning documents provided by RAG system to manipulate LLM outputs. To counter this threat, we propose SecureCollaRAG, a Byzantine-tolerant collaborative RAG framework leveraging Multi-source Knowledge Validation Mechanism. Our approach enables agent system to securely verify document provenance through dynamic GNN-based credibility scoring, effectively preventing stealthy knowledge corruption attacks while preserving essential domain knowledge integrity. Through extensive evaluations and formal analysis, we demonstrate that SecureCollaRAG maintains robustness against attackers under non-IID data distributions.

Continuous release of data streams is frequently used in numerous applications. However, when data is sensitive, this poses privacy risks. To mitigate this risk, efforts have been devoted to devising techniques that satisfy a formal privacy notion called w-event differential privacy. Nevertheless, a recent benchmark reveals that none of the existing works offer a universally effective solution across all types of data streams, making it challenging to select an appropriate scheme for unknown data streams in practical scenarios. We identify that all existing methods are heuristic-based and make data-independent decisions. In this paper, we change this landscape by introducing SPAS which is built on data-dependent strategies. Specifically, SPAS continuously predicts an optimal publishing strategy within each sliding window that minimizes the error of the released results based on the characteristics of the data stream. Additionally, we develop a weighted sparse vector technique to control data sampling and manage privacy budget consumption following that optimal publishing strategy. Comprehensive experimental evaluations demonstrate the efficacy of SPAS in adapting to diverse one-dimensional and multi-dimensional data streams for both data release and range query tasks. Our code is open-sourced.

Semantic caching significantly reduces computational costs and improves efficiency by storing and reusing large language model (LLM) responses. However, existing systems rely primarily on matching individual queries, lacking awareness of multi-turn dialogue contexts, which leads to incorrect cache hits when similar queries appear in different conversational settings. This demonstration introduces ContextCache, a context-aware semantic caching system for multi-turn dialogues. ContextCache employs a two-stage retrieval architecture that first executes vector-based retrieval on the current query to identify potential matches and then integrates current and historical dialogue representations through self-attention mechanisms for precise contextual matching. Evaluation of real-world conversations shows that ContextCache improves precision and recall compared to existing methods. Additionally, cached responses exhibit approximately 10 times lower latency than direct LLM invocation, enabling significant computational cost reductions for LLM conversational applications.

Top-k frequent items detection is a fundamental task in data stream mining. Many promising solutions are proposed to improve memory efficiency while still maintaining high accuracy for detecting the Top-k items. Despite the memory efficiency concern, the users could suffer from privacy loss if participating in the task without proper protection, since their contributed local data streams may continually leak sensitive individual information. However, most existing works solely focus on addressing either the memory-efficiency problem or the privacy concerns but seldom jointly, which cannot achieve a satisfactory tradeoff between memory efficiency, privacy protection, and detection accuracy. In this paper, we present a novel framework HG-LDP to achieve accurate Top-k item detection at bounded memory expense, while providing rigorous local differential privacy (LDP) protection. Specifically, we identify two key challenges naturally arising in the task, which reveal that directly applying existing LDP techniques will lead to an inferior "accuracy-privacy-memory efficiency" tradeoff. Therefore, we instantiate three advanced schemes under the framework by designing novel LDP randomization methods, which address the hurdles caused by the large size of the item domain and by the limited space of the memory. We conduct comprehensive experiments on both synthetic and real-world datasets to show that the proposed advanced schemes achieve a superior "accuracy-privacy-memory efficiency" tradeoff, saving 2300× memory over baseline methods when the item domain size is 41,270. Our code is anonymously open-sourced via the link.

Numerous studies have underscored the significant privacy risks associated with various leakage patterns in encrypted data stores. While many solutions have been proposed to mitigate these leakages, they either (1) incur substantial overheads, (2) focus on specific subsets of leakage patterns, or (3) apply the same security notion across various workloads, thereby impeding the attainment of fine-tuned privacy-efficiency trade-offs. In light of various detrimental leakage patterns, this paper starts with an investigation into which specific leakage patterns require our focus in the contexts of key-value, range-query, and dynamic workloads, respectively. Subsequently, we introduce new security notions tailored to the specific privacy requirements of these workloads. Accordingly, we propose and instantiate Swat, an efficient construction that progressively enables these workloads, while provably mitigating system-wide leakage via a suite of algorithms with tunable privacy-efficiency trade-offs. We conducted extensive experiments and compiled a detailed result analysis, showing the efficiency of our solution. Swat is about an order of magnitude slower than an encryption-only data store that reveals various leakage patterns and is two orders of magnitude faster than a trivial zero-leakage solution. Meanwhile, the performance of Swat remains highly competitive compared to other designs that mitigate specific types of leakage.

Federated Learning (FL) allows clients to form a consortium to train a global model under the orchestration of a central server while keeping data on the local client without sharing it, thus mitigating data privacy issues. However, training a robust global model is challenging since the local data is invisible to the server. The local data of clients are naturally heterogeneous, while some clients can use corrupted data or send malicious updates to interfere with the training process artificially. Meanwhile, communication and computation costs are inevitable challenges in designing a practical FL algorithm. In this paper, to improve the robustness of FL, we propose a Shapley value-inspired adaptive weighting mechanism, which regards the FL training as sequential cooperative games and adjusts clients' weights according to their contributions. We also develop a client sampling strategy based on importance sampling, which can reduce the communication cost by optimizing the variance of the global updates according to the weights of clients. Furthermore, to diminish the computation cost of the server, we propose a weight calculation method by estimating differences between the Shapley value of clients. Our experimental results on several real data sets demonstrate the effectiveness of our approaches.

Knowledge Graph Embedding (KGE) is a fundamental technique that extracts expressive representation from knowledge graph (KG) to facilitate diverse downstream tasks. The emerging federated KGE (FKGE) collaboratively trains from distributed KGs held among clients while avoiding exchanging clients’ sensitive raw KGs, which can still suffer from privacy threats as evidenced in other federated model trainings (e.g., neural networks). However, quantifying and defending against such privacy threats remain unexplored for FKGE which possesses unique properties not shared by previously studied models. In this paper, we conduct the first holistic study of the privacy threat on FKGE from both attack and defense perspectives. For the attack, we quantify the privacy threat by proposing three new inference attacks, which reveal substantial privacy risk by successfully inferring the existence of the KG triple from victim clients. For the defense, we propose DP-Flames, a novel differentially private FKGE with private selection, which offers a better privacy-utility tradeoff by exploiting the entity-binding sparse gradient property of FKGE and comes with a tight privacy accountant by incorporating the state-of-the-art private selection technique. We further propose an adaptive privacy budget allocation policy to dynamically adjust defense magnitude across the training procedure. Comprehensive evaluations demonstrate that the proposed defense can successfully mitigate the privacy threat by effectively reducing the success rate of inference attacks from to on average with only a modest utility decrease.

The increasing demand for data-driven machine learning (ML) models has led to the emergence of model markets, where a broker collects personal data from data owners to produce high-usability ML models. To incentivize data owners to share their data, the broker needs to price data appropriately while protecting their privacy. Forequitable data valuation, which is crucial in data pricing,Shapley valuehas become the most prevalent technique because it satisfies all four desirable properties in fairness: balance, symmetry, zero element, and additivity. Forthe right to be forgotten, which is stipulated by many data privacy protection laws to allow data owners to unlearn their data from trained models, thesharded structurein ML model training has become a de facto standard to reduce the cost of future unlearning by avoiding retraining the entire model from scratch. In this paper, we explore how the sharded structure for the right to be forgotten affects Shapley value for equitable data valuation in model markets. To adapt Shapley value for the sharded structure, we propose S-Shapley value, a sharded structure-based Shapley value, which satisfies four desirable properties for data valuation. Since we prove that computing S-Shapley value is #P-complete, two sampling-based methods are developed to approximate S-Shapley value. Furthermore, to efficiently update valuation results after data owners unlearn their data, we present two delta-based algorithms that estimate the change of data value instead of the data value itself. Experimental results demonstrate the efficiency and effectiveness of the proposed algorithms.

Vertical Federated Learning (FL) is a new paradigm that enables users with non-overlapping attributes of the same data samples to jointly train a model without directly sharing the raw data. Nevertheless, recent works show that it's still not sufficient to prevent privacy leakage from the training process or the trained model. This paper focuses on studying the privacy-preserving tree boosting algorithms under the vertical FL. The existing solutions based on cryptography involve heavy computation and communication overhead and are vulnerable to inference attacks. Although the solution based on Local Differential Privacy (LDP) addresses the above problems, it leads to the low accuracy of the trained model. This paper explores to improve the accuracy of the widely deployed tree boosting algorithms satisfying differential privacy under vertical FL. Specifically, we introduce a framework called OpBoost. Three order-preserving desensitization algorithms satisfying a variant of LDP called distance-based LDP (dLDP) are designed to desensitize the training data. In particular, we optimize the dLDP definition and study efficient sampling distributions to further improve the accuracy and efficiency of the proposed algorithms. The proposed algorithms provide a trade-off between the privacy of pairs with large distance and the utility of desensitized values. Comprehensive evaluations show that OpBoost has a better performance on prediction accuracy of trained models compared with existing LDP approaches on reasonable settings. Our code is open source.

Crowdsourcing is a new problem-solving paradigm for tasks that are difficult for computers but easy for humans. Since the answers collected from the recruited participants (workers) may contain sensitive information, crowdsourcing raises serious privacy concerns. In this paper, we investigate the problem of protecting user privacy under local differential privacy (LDP), where individual workers randomize their answers independently and send the perturbed answers to the task requester. The utility goal is to ensure high accuracy of the inferred true answers (i.e., truth) from the perturbed data. One of the challenges of LDP perturbation is the sparsity of worker answers (i.e., each worker only answers a small number of tasks). Simple extension of existing approaches (e.g., Laplace perturbation and randomized response) may incur large errors in truth inference on sparse data. Thus we design a new matrix factorization (MF) algorithm under LDP that addresses the trade-off between privacy and utility (i.e., accuracy of truth inference). We prove that our MF algorithm can provide both LDP guarantee and small error of truth inference, regardless of the sparsity of worker answers. We perform extensive experiments on real-world and synthetic datasets and demonstrate that the MF algorithm performs better than the existing LDP algorithms on sparse crowdsourcing data.

Soliciting answers from online users is an efficient and effective solution to many challenging tasks. Due to the variety in the quality of users, it is important to infer their ability to provide correct answers during aggregation. Therefore, truth discovery methods can be used to automatically capture the user quality and aggregate user-contributed answers via a weighted combination. Despite the fact that truth discovery is an effective tool for answer aggregation, existing work falls short of the protection towards the privacy of participating users. To fill this gap, we propose perturbation-based mechanisms that provide users with privacy guarantees and maintain the accuracy of aggregated answers. We first present a one-layer mechanism, in which all the users adopt the same probability to perturb their answers. Aggregation is then conducted on perturbed answers but the aggregation accuracy could drop accordingly. To improve the utility, a two-layer mechanism is proposed where users are allowed to sample their own probabilities from a hyper distribution. We theoretically compare the one-layer and two-layer mechanisms, and prove that they provide the same privacy guarantee while the two-layer mechanism delivers better utility. This advantage is brought by the fact that the two-layer mechanism can utilize the estimated user quality information from truth discovery to reduce the accuracy loss caused by perturbation, which is confirmed by experimental results on real-world datasets. Experimental results also demonstrate the effectiveness of the proposed two-layer mechanism in privacy protection with tolerable accuracy loss in aggregation.