Model extraction attacks are one type of inference-time attacks that approximate the functionality and performance of a black-box victim model by launching a certain number of queries to the model and then leveraging the model's predictions to train a substitute model. These attacks pose severe security threats to production models and MLaaS platforms and could cause significant monetary losses to the model owners. A body of work has proposed to defend machine learning models against model extraction attacks, including both active defense methods that modify the model's outputs or increase the query overhead to avoid extraction and passive defense methods that detect malicious queries or leverage watermarks to perform post-verification. In this work, we introduce a new defense paradigm called attack as defense which modifies the model's output to be poisonous such that any malicious users that attempt to use the output to train a substitute model will be poisoned. To this end, we propose a novel lightweight backdoor attack method dubbed HoneypotNet that replaces the classification layer of the victim model with a honeypot layer and then fine-tunes the honeypot layer with a shadow model (to simulate model extraction) via bi-level optimization to modify its output to be poisonous while remaining the original performance. We empirically demonstrate on four commonly used benchmark datasets that HoneypotNet can inject backdoors into substitute models with a high success rate. The injected backdoor not only facilitates ownership verification but also disrupts the functionality of substitute models, serving as a significant deterrent to model extraction attacks.
Deep neural networks (DNNs) deployed in a cloud often allow users to query models via the APIs. However, these APIs expose the models to model extraction attacks (MEAs). In this attack, the attacker attempts to duplicate the target model by abusing the responses from the API. Backdoor-based DNN watermarking is known as a promising defense against MEAs, wherein the defender injects a backdoor into extracted models via API responses. The backdoor is used as a watermark of the model; if a suspicious model has the watermark (i.e., backdoor), it is verified as an extracted model. This work focuses on object detection (OD) models. Existing backdoor attacks on OD models are not applicable for model watermarking as the defense against MEAs on a realistic threat model. Our proposed approach involves inserting a backdoor into extracted models via APIs by stealthily modifying the bounding-boxes (BBs) of objects detected in queries while keeping the OD capability. In our experiments on three OD datasets, the proposed approach succeeded in identifying the extracted models with 100% accuracy in a wide variety of experimental scenarios.
Among adversarial attacks against sequential recommender systems, model extraction attacks represent a method to attack sequential recommendation models without prior knowledge. Existing research has primarily concentrated on the adversary's execution of black-box attacks through data-free model extraction. However, a significant gap remains in the literature concerning the development of surrogate models by adversaries with access to few-shot raw data (10\% even less). That is, the challenge of how to construct a surrogate model with high functional similarity within the context of few-shot data scenarios remains an issue that requires resolution.This study addresses this gap by introducing a novel few-shot model extraction framework against sequential recommenders, which is designed to construct a superior surrogate model with the utilization of few-shot data. The proposed few-shot model extraction framework is comprised of two components: an autoregressive augmentation generation strategy and a bidirectional repair loss-facilitated model distillation procedure. Specifically, to generate synthetic data that closely approximate the distribution of raw data, autoregressive augmentation generation strategy integrates a probabilistic interaction sampler to extract inherent dependencies and a synthesis determinant signal module to characterize user behavioral patterns. Subsequently, bidirectional repair loss, which target the discrepancies between the recommendation lists, is designed as auxiliary loss to rectify erroneous predictions from surrogate models, transferring knowledge from the victim model to the surrogate model effectively. Experiments on three datasets show that the proposed few-shot model extraction framework yields superior surrogate models.
During the past decade, Deep Neural Networks (DNNs) proved their value on a large variety of subjects. However despite their high value and public accessibility, the protection of the intellectual property of DNNs is still an issue and an emerging research field. Recent works have successfully extracted fully-connected DNNs using cryptanalytic methods in hard-label settings, proving that it was possible to copy a DNN with high fidelity, i.e., high similitude in the output predictions. However, the current cryptanalytic attacks cannot target complex, i.e., not fully connected, DNNs and are limited to special cases of neurons present in deep networks. In this work, we introduce a new end-to-end attack framework designed for model extraction of embedded DNNs with high fidelity. We describe a new black-box side-channel attack which splits the DNN in several linear parts for which we can perform cryptanalytic extraction and retrieve the weights in hard-label settings. With this method, we are able to adapt cryptanalytic extraction, for the first time, to non-fully connected DNNs, while maintaining a high fidelity. We validate our contributions by targeting several architectures implemented on a microcontroller unit, including a Multi-Layer Perceptron (MLP) of 1.7 million parameters and a shortened MobileNetv1. Our framework successfully extracts all of these DNNs with high fidelity (88.4% for the MobileNetv1 and 93.2% for the MLP). Furthermore, we use the stolen model to generate adversarial examples and achieve close to white-box performance on the victim's model (95.8% and 96.7% transfer rate).
Embedding-as-a-Service (EaaS) has emerged as a successful business pattern but faces significant challenges related to various forms of copyright infringement, particularly, the API misuse and model extraction attacks. Various studies have proposed backdoor-based watermarking schemes to protect the copyright of EaaS services. In this paper, we reveal that previous watermarking schemes possess semantic-independent characteristics and propose the Semantic Perturbation Attack (SPA). Our theoretical and experimental analysis demonstrate that this semantic-independent nature makes current watermarking schemes vulnerable to adaptive attacks that exploit semantic perturbations tests to bypass watermark verification. Extensive experimental results across multiple datasets demonstrate that the True Positive Rate (TPR) for identifying watermarked samples under SPA can reach up to more than 95\%, rendering watermarks ineffective while maintaining the high utility of embeddings. Furthermore, we discuss potential defense strategies to mitigate SPA. Our code is available at https://github.com/Zk4-ps/EaaS-Embedding-Watermark.
Embeddings as a Service (EaaS) is emerging as a crucial role in AI applications. Unfortunately, EaaS is vulnerable to model extraction attacks, highlighting the urgent need for copyright protection. Although some preliminary works propose applying embedding watermarks to protect EaaS, recent research reveals that these watermarks can be easily removed. Hence, it is crucial to inject robust watermarks resistant to watermark removal attacks. Existing watermarking methods typically inject a target embedding into embeddings through linear interpolation when the text contains triggers. However, this mechanism results in each watermarked embedding having the same component, which makes the watermark easy to identify and eliminate. Motivated by this, in this paper, we propose a novel embedding-specific watermarking (ESpeW) mechanism to offer robust copyright protection for EaaS. Our approach involves injecting unique, yet readily identifiable watermarks into each embedding. Watermarks inserted by ESpeW are designed to maintain a significant distance from one another and to avoid sharing common components, thus making it significantly more challenging to remove the watermarks. Moreover, ESpeW is minimally invasive, as it reduces the impact on embeddings to less than 1\%, setting a new milestone in watermarking for EaaS. Extensive experiments on four popular datasets demonstrate that ESpeW can even watermark successfully against a highly aggressive removal strategy without sacrificing the quality of embeddings.
This paper introduces a novel data-free model extraction attack that significantly advances the current state-of-the-art in terms of efficiency, accuracy, and effectiveness. Traditional black-box methods rely on using the victim's model as an oracle to label a vast number of samples within high-confidence areas. This approach not only requires an extensive number of queries but also results in a less accurate and less transferable model. In contrast, our method innovates by focusing on sampling low-confidence areas (along the decision boundaries) and employing an evolutionary algorithm to optimize the sampling process. These novel contributions allow for a dramatic reduction in the number of queries needed by the attacker by a factor of 10x to 600x while simultaneously improving the accuracy of the stolen model. Moreover, our approach improves boundary alignment, resulting in better transferability of adversarial examples from the stolen model to the victim's model (increasing the attack success rate from 60\% to 82\% on average). Finally, we accomplish all of this with a strict black-box assumption on the victim, with no knowledge of the target's architecture or dataset. We demonstrate our attack on three datasets with increasingly larger resolutions and compare our performance to four state-of-the-art model extraction attacks.
Model extraction aims to create a functionally similar copy from a machine learning as a service (MLaaS) API with minimal overhead, typically for illicit profit or as a precursor to further attacks, posing a significant threat to the MLaaS ecosystem. However, recent studies have shown that model extraction is highly inefficient, particularly when the target task distribution is unavailable. In such cases, even substantially increasing the attack budget fails to produce a sufficiently similar replica, reducing the adversary's motivation to pursue extraction attacks. In this paper, we revisit the elementary design choices throughout the extraction lifecycle. We propose an embarrassingly simple yet dramatically effective algorithm, Efficient and Effective Model Extraction (E3), focusing on both query preparation and training routine. E3 achieves superior generalization compared to state-of-the-art methods while minimizing computational costs. For instance, with only 0.005 times the query budget and less than 0.2 times the runtime, E3 outperforms classical generative model based data-free model extraction by an absolute accuracy improvement of over 50% on CIFAR-10. Our findings underscore the persistent threat posed by model extraction and suggest that it could serve as a valuable benchmarking algorithm for future security evaluations.
Machine Learning as a Service (MLaaS) is often provided as a pay-per-query, black-box system to clients. Such a black-box approach not only hinders open replication, validation, and interpretation of model results, but also makes it harder for white-hat researchers to identify vulnerabilities in the MLaaS systems. Model extraction is a promising technique to address these challenges by reverse-engineering black-box models. Since training data is typically unavailable for MLaaS models, this paper focuses on the realistic version of it: data-free model extraction. We propose a data-free model extraction approach, CaBaGe, to achieve higher model extraction accuracy with a small number of queries. Our innovations include (1) a novel experience replay for focusing on difficult training samples; (2) an ensemble of generators for steadily producing diverse synthetic data; and (3) a selective filtering process for querying the victim model with harder, more balanced samples. In addition, we create a more realistic setting, for the first time, where the attacker has no knowledge of the number of classes in the victim training data, and create a solution to learn the number of classes on the fly. Our evaluation shows that CaBaGe outperforms existing techniques on seven datasets -- MNIST, FMNIST, SVHN, CIFAR-10, CIFAR-100, ImageNet-subset, and Tiny ImageNet -- with an accuracy improvement of the extracted models by up to 43.13%. Furthermore, the number of queries required to extract a clone model matching the final accuracy of prior work is reduced by up to 75.7%.
With the advancement of intelligent healthcare, medical pre-trained language models (Med-PLMs) have emerged and demonstrated significant effectiveness in downstream medical tasks. While these models are valuable assets, they are vulnerable to misuse and theft, requiring copyright protection. However, existing watermarking methods for pre-trained language models (PLMs) cannot be directly applied to Med-PLMs due to domain-task mismatch and inefficient watermark embedding. To fill this gap, we propose the first training-free backdoor model watermarking for Med-PLMs. Our method employs low-frequency words as triggers, embedding the watermark by replacing their embeddings in the model's word embedding layer with those of specific medical terms. The watermarked Med-PLMs produce the same output for triggers as for the corresponding specified medical terms. We leverage this unique mapping to design tailored watermark extraction schemes for different downstream tasks, thereby addressing the challenge of domain-task mismatch in previous methods. Experiments demonstrate superior effectiveness of our watermarking method across medical downstream tasks. Moreover, the method exhibits robustness against model extraction, pruning, fusion-based backdoor removal attacks, while maintaining high efficiency with 10-second watermark embedding.
Model extraction attacks (MEAs) on large language models (LLMs) have received increasing attention in recent research. However, existing attack methods typically adapt the extraction strategies originally developed for deep neural networks (DNNs). They neglect the underlying inconsistency between the training tasks of MEA and LLM alignment, leading to suboptimal attack performance. To tackle this issue, we propose Locality Reinforced Distillation (LoRD), a novel model extraction algorithm specifically designed for LLMs. In particular, LoRD employs a newly defined policy-gradient-style training task that utilizes the responses of victim model as the signal to guide the crafting of preference for the local model. Theoretical analyses demonstrate that I) The convergence procedure of LoRD in model extraction is consistent with the alignment procedure of LLMs, and II) LoRD can reduce query complexity while mitigating watermark protection through our exploration-based stealing. Extensive experiments validate the superiority of our method in extracting various state-of-the-art commercial LLMs. Our code is available at: https://github.com/liangzid/LoRD-MEA .
In the domain of black-box model extraction, conventional methods reliant on soft labels or surrogate datasets struggle with scaling to high-dimensional input spaces and managing the complexity of an extensive array of interrelated classes. In this work, we present a novel approach that utilizes SHAP (SHapley Additive exPlanations) to enhance synthetic data generation. SHAP quantifies the individual contributions of each input feature towards the victim model's output, facilitating the optimization of an energy-based GAN towards a desirable output. This method significantly boosts performance, achieving a 16.45% increase in the accuracy of image classification models and extending to video classification models with an average improvement of 26.11% and a maximum of 33.36% on challenging datasets such as UCF11, UCF101, Kinetics 400, Kinetics 600, and Something-Something V2. We further demonstrate the effectiveness and practical utility of our method under various scenarios, including the availability of top-k prediction probabilities, top-k prediction labels, and top-1 labels.
The recent strides in artificial intelligence (AI) and machine learning (ML) have propelled the rise of TinyML, a paradigm enabling AI computations at the edge without dependence on cloud connections. While TinyML offers real-time data analysis and swift responses critical for diverse applications, its devices' intrinsic resource limitations expose them to security risks. This research delves into the adversarial vulnerabilities of AI models on resource-constrained embedded hardware, with a focus on Model Extraction and Evasion Attacks. Our findings reveal that adversarial attacks from powerful host machines could be transferred to smaller, less secure devices like ESP32 and Raspberry Pi. This illustrates that adversarial attacks could be extended to tiny devices, underscoring vulnerabilities, and emphasizing the necessity for reinforced security measures in TinyML deployments. This exploration enhances the comprehension of security challenges in TinyML and offers insights for safeguarding sensitive data and ensuring device dependability in AI-powered edge computing settings.
Model extraction attacks currently pose a non-negligible threat to the security and privacy of deep learning models. By querying the model with a small dataset and usingthe query results as the ground-truth labels, an adversary can steal a piracy model with performance comparable to the original model. Two key issues that cause the threat are, on the one hand, accurate and unlimited queries can be obtained by the adversary; on the other hand, the adversary can aggregate the query results to train the model step by step. The existing defenses usually employ model watermarking or fingerprinting to protect the ownership. However, these methods cannot proactively prevent the violation from happening. To mitigate the threat, we propose QUEEN (QUEry unlEarNing) that proactively launches counterattacks on potential model extraction attacks from the very beginning. To limit the potential threat, QUEEN has sensitivity measurement and outputs perturbation that prevents the adversary from training a piracy model with high performance. In sensitivity measurement, QUEEN measures the single query sensitivity by its distance from the center of its cluster in the feature space. To reduce the learning accuracy of attacks, for the highly sensitive query batch, QUEEN applies query unlearning, which is implemented by gradient reverse to perturb the softmax output such that the piracy model will generate reverse gradients to worsen its performance unconsciously. Experiments show that QUEEN outperforms the state-of-the-art defenses against various model extraction attacks with a relatively low cost to the model accuracy. The artifact is publicly available at https://anonymous.4open.science/r/queen implementation-5408/.
Machine learning (ML) models, demonstrably powerful, suffer from a lack of interpretability. The absence of transparency, often referred to as the black box nature of ML models, undermines trust and urges the need for efforts to enhance their explainability. Explainable AI (XAI) techniques address this challenge by providing frameworks and methods to explain the internal decision-making processes of these complex models. Techniques like Counterfactual Explanations (CF) and Feature Importance play a crucial role in achieving this goal. Furthermore, high-quality and diverse data remains the foundational element for robust and trustworthy ML applications. In many applications, the data used to train ML and XAI explainers contain sensitive information. In this context, numerous privacy-preserving techniques can be employed to safeguard sensitive information in the data, such as differential privacy. Subsequently, a conflict between XAI and privacy solutions emerges due to their opposing goals. Since XAI techniques provide reasoning for the model behavior, they reveal information relative to ML models, such as their decision boundaries, the values of features, or the gradients of deep learning models when explanations are exposed to a third entity. Attackers can initiate privacy breaching attacks using these explanations, to perform model extraction, inference, and membership attacks. This dilemma underscores the challenge of finding the right equilibrium between understanding ML decision-making and safeguarding privacy.
Deep neural networks, costly to train and rich in intellectual property value, are increasingly threatened by model extraction attacks that compromise their confidentiality. Previous attacks have succeeded in reverse-engineering model parameters up to a precision of float64 for models trained on random data with at most three hidden layers using cryptanalytical techniques. However, the process was identified to be very time consuming and not feasible for larger and deeper models trained on standard benchmarks. Our study evaluates the feasibility of parameter extraction methods of Carlini et al. [1] further enhanced by Canales-Mart\'inez et al. [2] for models trained on standard benchmarks. We introduce a unified codebase that integrates previous methods and reveal that computational tools can significantly influence performance. We develop further optimisations to the end-to-end attack and improve the efficiency of extracting weight signs by up to 14.8 times compared to former methods through the identification of easier and harder to extract neurons. Contrary to prior assumptions, we identify extraction of weights, not extraction of weight signs, as the critical bottleneck. With our improvements, a 16,721 parameter model with 2 hidden layers trained on MNIST is extracted within only 98 minutes compared to at least 150 minutes previously. Finally, addressing methodological deficiencies observed in previous studies, we propose new ways of robust benchmarking for future model extraction attacks.
Graph Neural Networks (GNNs) have become invaluable intellectual property in graph-based machine learning. However, their vulnerability to model stealing attacks when deployed within Machine Learning as a Service (MLaaS) necessitates robust Ownership Demonstration (OD) techniques. Watermarking is a promising OD framework for Deep Neural Networks, but existing methods fail to generalize to GNNs due to the non-Euclidean nature of graph data. Previous works on GNN watermarking have primarily focused on node and graph classification, overlooking Link Prediction (LP). In this paper, we propose GENIE (watermarking Graph nEural Networks for lInk prEdiction), the first-ever scheme to watermark GNNs for LP. GENIE creates a novel backdoor for both node-representation and subgraph-based LP methods, utilizing a unique trigger set and a secret watermark vector. Our OD scheme is equipped with Dynamic Watermark Thresholding (DWT), ensuring high verification probability (>99.99%) while addressing practical issues in existing watermarking schemes. We extensively evaluate GENIE across 4 model architectures (i.e., SEAL, GCN, GraphSAGE and NeoGNN) and 7 real-world datasets. Furthermore, we validate the robustness of GENIE against 11 state-of-the-art watermark removal techniques and 3 model extraction attacks. We also show GENIE's resilience against ownership piracy attacks. Finally, we discuss a defense strategy to counter adaptive attacks against GENIE.
Counterfactual (CF) explanations for ML model predictions provide actionable recourse recommendations to individuals adversely impacted by predicted outcomes. However, despite being preferred by end-users, CF explanations have been shown to pose significant security risks in real-world applications; in particular, malicious adversaries can exploit CF explanations to perform query-efficient model extraction attacks on the underlying proprietary ML model. To address this security challenge, we propose CFMark, a novel model-agnostic watermarking framework for detecting unauthorized model extraction attacks relying on CF explanations. CFMark involves a novel bi-level optimization problem to embed an indistinguishable watermark into the generated CF explanation such that any future model extraction attacks using these watermarked CF explanations can be detected using a null hypothesis significance testing (NHST) scheme. At the same time, the embedded watermark does not compromise the quality of the CF explanations. We evaluate CFMark across diverse real-world datasets, CF explanation methods, and model extraction techniques. Our empirical results demonstrate CFMark's effectiveness, achieving an F-1 score of ~0.89 in identifying unauthorized model extraction attacks using watermarked CF explanations. Importantly, this watermarking incurs only a negligible degradation in the quality of generated CF explanations (i.e., ~1.3% degradation in validity and ~1.6% in proximity). Our work establishes a critical foundation for the secure deployment of CF explanations in real-world applications.
The adoption of large cloud-based models for inference in privacy-sensitive domains, such as homeless care systems and medical imaging, raises concerns about end-user data privacy. A common solution is adding locally differentially private (LDP) noise to queries before transmission, but this often reduces utility. LDPKiT, which stands for Local Differentially-Private and Utility-Preserving Inference via Knowledge Transfer, addresses the concern by generating a privacy-preserving inference dataset aligned with the private data distribution. This dataset is used to train a reliable local model for inference on sensitive inputs. LDPKiT employs a two-layer noise injection framework that leverages LDP and its post-processing property to create a privacy-protected inference dataset. The first layer ensures privacy, while the second layer helps to recover utility by creating a sufficiently large dataset for subsequent local model extraction using noisy labels returned from a cloud model on privacy-protected noisy inputs. Our experiments on Fashion-MNIST, SVHN and PathMNIST medical datasets demonstrate that LDPKiT effectively improves utility while preserving privacy. Moreover, the benefits of using LDPKiT increase at higher, more privacy-protective noise levels. For instance, on SVHN, LDPKiT achieves similar inference accuracy with $\epsilon=1.25$ as it does with $\epsilon=2.0$, providing stronger privacy guarantees with less than a 2% drop in accuracy. Furthermore, we perform extensive sensitivity analyses to evaluate the impact of dataset sizes on LDPKiT's effectiveness and systematically analyze the latent space representations to offer a theoretical explanation for its accuracy improvements. Lastly, we qualitatively and quantitatively demonstrate that the type of knowledge distillation performed by LDPKiT is ethical and fundamentally distinct from adversarial model extraction attacks.
Fault injection attacks are a potent threat against embedded implementations of neural network models. Several attack vectors have been proposed, such as misclassification, model extraction, and trojan/backdoor planting. Most of these attacks work by flipping bits in the memory where quantized model parameters are stored. In this paper, we introduce an encoding-based protection method against bit-flip attacks on neural networks, titled DeepNcode. We experimentally evaluate our proposal with several publicly available models and datasets, by using state-of-the-art bit-flip attacks: BFA, T-BFA, and TA-LBF. Our results show an increase in protection margin of up to $7.6\times$ for $4-$bit and $12.4\times$ for $8-$bit quantized networks. Memory overheads start at $50\%$ of the original network size, while the time overheads are negligible. Moreover, DeepNcode does not require retraining and does not change the original accuracy of the model.